Skip to content

GitLab

Commit 2a768425 authored by Mateusz Ozga's avatar Mateusz Ozga
Browse files

Merge remote-tracking branch 'origin/develop' into mozga-amd/universal_gemm_weight

parents f358055f 1c45ca35
Hide whitespace changes
Inline Side-by-side
Showing with 3431 additions and 2082 deletions
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp
View file @ 2a768425
......@@ -729,6 +729,7 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
return str.str();
}
REGISTER_EXTRA_PRINTING_METHODS
};
} // namespace device
......
include/ck/utility/amd_ck_fp8.hpp
View file @ 2a768425
......@@ -18,6 +18,20 @@
#define CK_USE_OCP_FP8 0
#endif
namespace {
// https://en.cppreference.com/w/cpp/types/conditional
template <bool B, class T, class F>
struct conditional
{
using type = T;
};
template <class T, class F>
struct conditional<false, T, F>
{
using type = F;
};
} // namespace
namespace ck {
using f8_fnuz_t = _BitInt(8);
......@@ -191,11 +205,10 @@ __host__ __device__ static inline T cast_from_f8(fp8_storage_t x)
}
}
typename __hip_internal::conditional<
typename conditional<
sizeof(T) == 2,
unsigned short int,
typename __hip_internal::conditional<sizeof(T) == 4, unsigned int, unsigned long long>::
type>::type retval;
typename conditional<sizeof(T) == 4, unsigned int, unsigned long long>::type>::type retval;
if constexpr(we == 5 && is_half && !is_fnuz)
{
......@@ -538,11 +551,10 @@ __host__ __device__ static inline fp8_storage_t cast_to_f8(T _x, unsigned int rn
constexpr int mfmt = (sizeof(T) == 8) ? 52 : ((sizeof(T) == 4) ? 23 : 10);
using T_bitwise = typename __hip_internal::conditional<
using T_bitwise = typename conditional<
sizeof(T) == 2,
unsigned short int,
typename __hip_internal::conditional<sizeof(T) == 4, unsigned int, unsigned long long>::
type>::type;
typename conditional<sizeof(T) == 4, unsigned int, unsigned long long>::type>::type;
T_bitwise x_bitwise = bit_cast<T_bitwise>(_x);
unsigned long long x{x_bitwise};
......
include/ck_tile/core/arch/amd_buffer_addressing.hpp
View file @ 2a768425
......@@ -1303,8 +1303,8 @@ CK_TILE_DEVICE thread_buffer<T, N> amd_buffer_load_impl(int32x4_t src_wave_buffe
static_assert(
(std::is_same<T, double>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
(std::is_same<T, float>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(std::is_same<T, fp16_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(std::is_same<T, bf16_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(std::is_same<T, fp16_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
(std::is_same<T, bf16_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
(std::is_same<T, int32_t>::value &&
(N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(std::is_same<T, fp8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
......
include/ck_tile/core/container/meta_data_buffer.hpp
View file @ 2a768425
......@@ -30,7 +30,7 @@ struct meta_data_buffer
{
constexpr index_t size = sizeof(T);
auto tmp = bit_cast<array<std::byte, size>>(data);
auto tmp = ck_tile::bit_cast<array<std::byte, size>>(data);
for(int i = 0; i < size; i++)
{
......@@ -66,7 +66,7 @@ struct meta_data_buffer
pos++;
}
data = bit_cast<T>(tmp);
data = ck_tile::bit_cast<T>(tmp);
}
return data;
......@@ -86,7 +86,7 @@ struct meta_data_buffer
pos++;
}
auto data = bit_cast<T>(tmp);
auto data = ck_tile::bit_cast<T>(tmp);
return data;
}
......
include/ck_tile/core/tensor/static_distributed_tensor.hpp
View file @ 2a768425
......@@ -29,6 +29,7 @@ struct static_distributed_tensor
remove_cvref_t<decltype(StaticTileDistribution{}.get_ys_to_d_descriptor())>;
static constexpr index_t kThreadElementSpaceSize = ThreadTensorDesc{}.get_element_space_size();
static_assert(0 < kThreadElementSpaceSize, "Make sure tile distribution is valid");
CK_TILE_HOST_DEVICE static constexpr auto get_num_of_dimension()
{
......
include/ck_tile/host/arg_parser.hpp
View file @ 2a768425
......@@ -15,11 +15,14 @@
namespace ck_tile {
/*
* a host side utility, arg parser for
* -[key0]=[value0] -[key1]=[value1] ...
* a host side utility, arg parser for, either
* -[key0] = [value0, value1, value2]
* or
* -[key0]=[value0] -[key1]=[value1] ...
*/
class ArgParser
{
public:
class Arg
{
......@@ -187,6 +190,45 @@ class ArgParser
return value;
}
std::vector<std::string> get_string_vec(const std::string& name,
const std::string& delimiter = ",") const
{
if(get_str(name).empty())
{
return {};
}
std::string s = get_str(name);
std::vector<std::string> tokens;
size_t pos = 0;
std::string token;
while((pos = s.find(delimiter)) != std::string::npos)
{
token = s.substr(0, pos);
tokens.push_back(token);
s.erase(0, pos + delimiter.length());
}
tokens.push_back(s);
return tokens;
}
std::vector<int> get_int_vec(const std::string& name, const std::string& delimiter = ",") const
{
if(get_str(name).empty())
{
return {};
}
const std::vector<std::string> args = get_string_vec(name, delimiter);
std::vector<int> tokens;
tokens.reserve(static_cast<int>(args.size()));
for(const std::string& token : args)
{
int value = atoi(token.c_str());
tokens.push_back(value);
}
return tokens;
}
private:
std::unordered_map<std::string, Arg> input_map;
std::vector<std::string> keys;
......
include/ck_tile/host/reference/reference_gemm.hpp
View file @ 2a768425
......@@ -97,9 +97,9 @@ template <typename ADataType,
typename LayoutA,
typename LayoutB,
typename LayoutC>
void reference_gemm_gpu(DeviceMem& a_device,
DeviceMem& b_device,
DeviceMem& c_device,
void reference_gemm_gpu(ADataType* a_ptr,
BDataType* b_ptr,
CDataType* c_ptr,
index_t M,
index_t N,
index_t K,
......@@ -107,79 +107,13 @@ void reference_gemm_gpu(DeviceMem& a_device,
index_t stride_b,
index_t stride_c)
{
ADataType* d_A;
BDataType* d_B;
CDataType* d_C;
hipError_t errA = hipMalloc(&d_A, M * K * sizeof(ADataType));
hipError_t errB = hipMalloc(&d_B, N * K * sizeof(BDataType));
hipError_t errC = hipMalloc(&d_C, M * N * sizeof(CDataType));
if(errA != hipSuccess)
{
std::cerr << "Error allocating device memory for A: " << hipGetErrorString(errA)
<< std::endl;
return; // Early exit on error
}
if(errB != hipSuccess)
{
std::cerr << "Error allocating device memory for B: " << hipGetErrorString(errB)
<< std::endl;
return; // Early exit on error
}
if(errC != hipSuccess)
{
std::cerr << "Error allocating device memory for C: " << hipGetErrorString(errC)
<< std::endl;
return; // Early exit on error
}
errA = hipMemcpy(
d_A, a_device.GetDeviceBuffer(), M * K * sizeof(ADataType), hipMemcpyHostToDevice);
if(errA != hipSuccess)
{
std::cerr << "Error copying A to device: " << hipGetErrorString(errA) << std::endl;
}
errB = hipMemcpy(
d_B, b_device.GetDeviceBuffer(), N * K * sizeof(BDataType), hipMemcpyHostToDevice);
if(errB != hipSuccess)
{
std::cerr << "Error copying B to device: " << hipGetErrorString(errB) << std::endl;
}
int totalElements = M * N;
int numThreadsPerBlock = 256; // Common choice for threads per block
int numBlocks = (totalElements + numThreadsPerBlock - 1) / numThreadsPerBlock;
naive_gemm_kernel<ADataType, BDataType, AccDataType, CDataType, LayoutA, LayoutB, LayoutC>
<<<numBlocks, numThreadsPerBlock>>>(d_A, d_B, d_C, M, N, K, stride_a, stride_b, stride_c);
errC = hipMemcpy(
c_device.GetDeviceBuffer(), d_C, M * N * sizeof(CDataType), hipMemcpyDeviceToHost);
if(errC != hipSuccess)
{
std::cerr << "Error copying C to device: " << hipGetErrorString(errC) << std::endl;
}
errA = hipFree(d_A);
if(errA != hipSuccess)
{
std::cerr << "Error free the A memory: " << hipGetErrorString(errA) << std::endl;
}
errB = hipFree(d_B);
if(errB != hipSuccess)
{
std::cerr << "Error free the B memory: " << hipGetErrorString(errB) << std::endl;
}
errC = hipFree(d_C);
if(errC != hipSuccess)
{
std::cerr << "Error free the C memory: " << hipGetErrorString(errC) << std::endl;
}
<<<numBlocks, numThreadsPerBlock>>>(
a_ptr, b_ptr, c_ptr, M, N, K, stride_a, stride_b, stride_c);
return;
}
......@@ -191,9 +125,9 @@ template <typename ADataType,
typename LayoutA,
typename LayoutB,
typename LayoutC>
void reference_batched_gemm_gpu(DeviceMem& a_device,
DeviceMem& b_device,
DeviceMem& c_device,
void reference_batched_gemm_gpu(ADataType* a_ptr,
BDataType* b_ptr,
CDataType* c_ptr,
index_t M,
index_t N,
index_t K,
......@@ -205,94 +139,20 @@ void reference_batched_gemm_gpu(DeviceMem& a_device,
index_t batch_stride_C,
index_t batch_count)
{
ADataType* d_A;
BDataType* d_B;
CDataType* d_C;
hipError_t errA = hipMalloc(&d_A, batch_count * M * K * sizeof(ADataType));
hipError_t errB = hipMalloc(&d_B, batch_count * N * K * sizeof(BDataType));
hipError_t errC = hipMalloc(&d_C, batch_count * M * N * sizeof(CDataType));
if(errA != hipSuccess)
{
std::cerr << "Error allocating device memory for A: " << hipGetErrorString(errA)
<< std::endl;
return; // Early exit on error
}
if(errB != hipSuccess)
{
std::cerr << "Error allocating device memory for B: " << hipGetErrorString(errB)
<< std::endl;
return; // Early exit on error
}
if(errC != hipSuccess)
{
std::cerr << "Error allocating device memory for C: " << hipGetErrorString(errC)
<< std::endl;
return; // Early exit on error
}
errA = hipMemcpy(d_A,
a_device.GetDeviceBuffer(),
batch_count * M * K * sizeof(ADataType),
hipMemcpyHostToDevice);
if(errA != hipSuccess)
{
std::cerr << "Error copying A to device: " << hipGetErrorString(errA) << std::endl;
}
errB = hipMemcpy(d_B,
b_device.GetDeviceBuffer(),
batch_count * N * K * sizeof(BDataType),
hipMemcpyHostToDevice);
if(errB != hipSuccess)
{
std::cerr << "Error copying B to device: " << hipGetErrorString(errB) << std::endl;
}
int totalElements = M * N;
int numThreadsPerBlock = 256; // Common choice for threads per block
int numBlocks = (totalElements + numThreadsPerBlock - 1) / numThreadsPerBlock;
for(index_t batch_id = 0; batch_id < batch_count; ++batch_id)
{
ADataType* d_ATemp = d_A + batch_id * batch_stride_A;
BDataType* d_BTemp = d_B + batch_id * batch_stride_B;
CDataType* d_CTemp = d_C + batch_id * batch_stride_C;
ADataType* d_ATemp = a_ptr + batch_id * batch_stride_A;
BDataType* d_BTemp = b_ptr + batch_id * batch_stride_B;
CDataType* d_CTemp = c_ptr + batch_id * batch_stride_C;
naive_gemm_kernel<ADataType, BDataType, AccDataType, CDataType, LayoutA, LayoutB, LayoutC>
<<<numBlocks, numThreadsPerBlock>>>(
d_ATemp, d_BTemp, d_CTemp, M, N, K, stride_a, stride_b, stride_c);
}
errC = hipMemcpy(c_device.GetDeviceBuffer(),
d_C,
batch_count * M * N * sizeof(CDataType),
hipMemcpyDeviceToHost);
if(errC != hipSuccess)
{
std::cerr << "Error copying C to device: " << hipGetErrorString(errC) << std::endl;
}
errA = hipFree(d_A);
if(errA != hipSuccess)
{
std::cerr << "Error free the A memory: " << hipGetErrorString(errA) << std::endl;
}
errB = hipFree(d_B);
if(errB != hipSuccess)
{
std::cerr << "Error free the B memory: " << hipGetErrorString(errB) << std::endl;
}
errC = hipFree(d_C);
if(errC != hipSuccess)
{
std::cerr << "Error free the C memory: " << hipGetErrorString(errC) << std::endl;
}
return;
}
} // namespace ck_tile
include/ck_tile/ops/flatmm/block/uk/flatmm_sn_uk_gfx9_32x128x512_1x4x1_16x16x16.inc
View file @ 2a768425
......@@ -3,610 +3,815 @@
#endif
#if CK_TILE_FLATMM_UK_MFMA == CK_TILE_FLATMM_UK_MFMA_BF16
# define _UK_MFMA_ "v_mfma_f32_16x16x16_bf16"
#define _UK_MFMA_ "v_mfma_f32_16x16x16_bf16"
# define _UK_PK_CVT_(x0_, x1_, y_) \
" v_cmp_u_f32 s[36:37], " x0_ ", " x0_ " \n" \
" v_add3_u32 v50, " x0_ ", %[v_nan_lo], 1 \n" \
" v_cndmask_b32 v54, v50, %[v_nan_hi], s[36:37] \n" \
" v_cmp_u_f32 s[36:37], " x1_ ", " x1_ " \n" \
" v_add3_u32 v50, " x1_ ", %[v_nan_lo], 1 \n" \
" v_cndmask_b32 v55, v50, %[v_nan_hi], s[36:37] \n" \
#define _UK_PK_CVT_(x0_, x1_, y_) \
" v_cmp_u_f32 s[36:37], " x0_ ", " x0_ " \n" \
" v_add3_u32 v50, " x0_ ", %[v_nan_lo], 1 \n" \
" v_cndmask_b32 v54, v50, %[v_nan_hi], s[36:37] \n" \
" v_cmp_u_f32 s[36:37], " x1_ ", " x1_ " \n" \
" v_add3_u32 v50, " x1_ ", %[v_nan_lo], 1 \n" \
" v_cndmask_b32 v55, v50, %[v_nan_hi], s[36:37] \n" \
" v_perm_b32 " y_ ", v55, v54, s52 \n"
# define _UK_ATOMIC_ADD_ "global_atomic_pk_add_bf16"
#define _UK_ATOMIC_ADD_ "global_atomic_pk_add_bf16"
#elif CK_TILE_FLATMM_UK_MFMA == CK_TILE_FLATMM_UK_MFMA_FP16
#define _UK_MFMA_ "v_mfma_f32_16x16x16_f16"
# define _UK_PK_CVT_(x0_, x1_, y_) \
" v_cvt_f16_f32 v54, " x0_ " \n" \
" v_cvt_f16_f32 v55, " x1_ " \n" \
#define _UK_PK_CVT_(x0_, x1_, y_) \
" v_cvt_f16_f32 v54, " x0_ " \n" \
" v_cvt_f16_f32 v55, " x1_ " \n" \
" v_pack_b32_f16 " y_ ", v54, v55 \n"
# define _UK_ATOMIC_ADD_ "global_atomic_pk_add_f16"
#define _UK_ATOMIC_ADD_ "global_atomic_pk_add_f16"
#endif
";-------------------------------------------------------------\n"
" s_mov_b32 s52, 0x07060302 ; v_perm\n"
" s_mov_b64 s[38:39], exec ; save current exec\n"
" s_mov_b32 s8, %[s_res_o0] \n"
" s_mov_b32 s9, %[s_res_o1] \n"
" s_mov_b32 s12, %[s_res_b0] \n"
" s_mov_b32 s13, %[s_res_b1] \n"
" s_mov_b32 s14, %[s_res_b2] \n"
" s_mov_b32 s15, %[s_res_b3] \n"
" ds_read_b64 v[128:129], %[v_sld_y_os] offset:0 + %[sld_a_base] \n"
" ds_read_b64 v[130:131], %[v_sld_y_os] offset:128 + %[sld_a_base] \n"
" ds_read_b64 v[132:133], %[v_sld_y_os] offset:1024 + %[sld_a_base] \n"
" ds_read_b64 v[134:135], %[v_sld_y_os] offset:1152 + %[sld_a_base] \n"
" ds_read_b64 v[136:137], %[v_sld_y_os] offset:2048 + %[sld_a_base] \n"
" ds_read_b64 v[138:139], %[v_sld_y_os] offset:2176 + %[sld_a_base] \n"
" ds_read_b64 v[140:141], %[v_sld_y_os] offset:3072 + %[sld_a_base] \n"
" ds_read_b64 v[142:143], %[v_sld_y_os] offset:3200 + %[sld_a_base] \n"
" ds_read_b64 v[144:145], %[v_sld_y_os] offset:4096 + %[sld_a_base] \n"
" ds_read_b64 v[146:147], %[v_sld_y_os] offset:4224 + %[sld_a_base] \n"
" ds_read_b64 v[148:149], %[v_sld_y_os] offset:5120 + %[sld_a_base] \n"
" ds_read_b64 v[150:151], %[v_sld_y_os] offset:5248 + %[sld_a_base] \n"
" ds_read_b64 v[152:153], %[v_sld_y_os] offset:6144 + %[sld_a_base] \n"
" ds_read_b64 v[154:155], %[v_sld_y_os] offset:6272 + %[sld_a_base] \n"
" ds_read_b64 v[156:157], %[v_sld_y_os] offset:7168 + %[sld_a_base] \n"
" ds_read_b64 v[158:159], %[v_sld_y_os] offset:7296 + %[sld_a_base] \n"
" ds_read_b64 v[160:161], %[v_sld_y_os] offset:8192 + %[sld_a_base] \n"
" ds_read_b64 v[162:163], %[v_sld_y_os] offset:8320 + %[sld_a_base] \n"
" ds_read_b64 v[164:165], %[v_sld_y_os] offset:9216 + %[sld_a_base] \n"
" ds_read_b64 v[166:167], %[v_sld_y_os] offset:9344 + %[sld_a_base] \n"
" ds_read_b64 v[168:169], %[v_sld_y_os] offset:10240 + %[sld_a_base] \n"
" ds_read_b64 v[170:171], %[v_sld_y_os] offset:10368 + %[sld_a_base] \n"
" ds_read_b64 v[172:173], %[v_sld_y_os] offset:11264 + %[sld_a_base] \n"
" ds_read_b64 v[174:175], %[v_sld_y_os] offset:11392 + %[sld_a_base] \n"
" ds_read_b64 v[176:177], %[v_sld_y_os] offset:12288 + %[sld_a_base] \n"
" ds_read_b64 v[178:179], %[v_sld_y_os] offset:12416 + %[sld_a_base] \n"
" ds_read_b64 v[180:181], %[v_sld_y_os] offset:13312 + %[sld_a_base] \n"
" ds_read_b64 v[182:183], %[v_sld_y_os] offset:13440 + %[sld_a_base] \n"
" ds_read_b64 v[184:185], %[v_sld_y_os] offset:14336 + %[sld_a_base] \n"
" ds_read_b64 v[186:187], %[v_sld_y_os] offset:14464 + %[sld_a_base] \n"
" ds_read_b64 v[188:189], %[v_sld_y_os] offset:15360 + %[sld_a_base] \n"
" ds_read_b64 v[190:191], %[v_sld_y_os] offset:15488 + %[sld_a_base] \n"
" ds_read_b64 v[192:193], %[v_sld_y_os] offset:16384 + %[sld_a_base] \n"
" ds_read_b64 v[194:195], %[v_sld_y_os] offset:16512 + %[sld_a_base] \n"
" ds_read_b64 v[196:197], %[v_sld_y_os] offset:17408 + %[sld_a_base] \n"
" ds_read_b64 v[198:199], %[v_sld_y_os] offset:17536 + %[sld_a_base] \n"
" ds_read_b64 v[200:201], %[v_sld_y_os] offset:18432 + %[sld_a_base] \n"
" ds_read_b64 v[202:203], %[v_sld_y_os] offset:18560 + %[sld_a_base] \n"
" ds_read_b64 v[204:205], %[v_sld_y_os] offset:19456 + %[sld_a_base] \n"
" ds_read_b64 v[206:207], %[v_sld_y_os] offset:19584 + %[sld_a_base] \n"
" ds_read_b64 v[208:209], %[v_sld_y_os] offset:20480 + %[sld_a_base] \n"
" ds_read_b64 v[210:211], %[v_sld_y_os] offset:20608 + %[sld_a_base] \n"
" ds_read_b64 v[212:213], %[v_sld_y_os] offset:21504 + %[sld_a_base] \n"
" ds_read_b64 v[214:215], %[v_sld_y_os] offset:21632 + %[sld_a_base] \n"
" ds_read_b64 v[216:217], %[v_sld_y_os] offset:22528 + %[sld_a_base] \n"
" ds_read_b64 v[218:219], %[v_sld_y_os] offset:22656 + %[sld_a_base] \n"
" ds_read_b64 v[220:221], %[v_sld_y_os] offset:23552 + %[sld_a_base] \n"
" ds_read_b64 v[222:223], %[v_sld_y_os] offset:23680 + %[sld_a_base] \n"
" ds_read_b64 v[224:225], %[v_sld_y_os] offset:24576 + %[sld_a_base] \n"
" ds_read_b64 v[226:227], %[v_sld_y_os] offset:24704 + %[sld_a_base] \n"
" ds_read_b64 v[228:229], %[v_sld_y_os] offset:25600 + %[sld_a_base] \n"
" ds_read_b64 v[230:231], %[v_sld_y_os] offset:25728 + %[sld_a_base] \n"
" ds_read_b64 v[232:233], %[v_sld_y_os] offset:26624 + %[sld_a_base] \n"
" ds_read_b64 v[234:235], %[v_sld_y_os] offset:26752 + %[sld_a_base] \n"
" ds_read_b64 v[236:237], %[v_sld_y_os] offset:27648 + %[sld_a_base] \n"
" ds_read_b64 v[238:239], %[v_sld_y_os] offset:27776 + %[sld_a_base] \n"
" ds_read_b64 v[240:241], %[v_sld_y_os] offset:28672 + %[sld_a_base] \n"
" ds_read_b64 v[242:243], %[v_sld_y_os] offset:28800 + %[sld_a_base] \n"
" ds_read_b64 v[244:245], %[v_sld_y_os] offset:29696 + %[sld_a_base] \n"
" ds_read_b64 v[246:247], %[v_sld_y_os] offset:29824 + %[sld_a_base] \n"
" ds_read_b64 v[248:249], %[v_sld_y_os] offset:30720 + %[sld_a_base] \n"
" ds_read_b64 v[250:251], %[v_sld_y_os] offset:30848 + %[sld_a_base] \n"
" ds_read_b64 v[252:253], %[v_sld_y_os] offset:31744 + %[sld_a_base] \n"
" ds_read_b64 v[254:255], %[v_sld_y_os] offset:31872 + %[sld_a_base] \n"
" s_waitcnt 0 \n"
" buffer_load_dwordx4 acc[0:3], %[v_os_b0], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[4:7], %[v_os_b0], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[8:11], %[v_os_b0], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[12:15], %[v_os_b0], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[16:19], %[v_os_b1], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[20:23], %[v_os_b1], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[24:27], %[v_os_b1], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[28:31], %[v_os_b1], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[32:35], %[v_os_b2], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[36:39], %[v_os_b2], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[40:43], %[v_os_b2], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[44:47], %[v_os_b2], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[48:51], %[v_os_b3], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[52:55], %[v_os_b3], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[56:59], %[v_os_b3], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[60:63], %[v_os_b3], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[64:67], %[v_os_b4], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[68:71], %[v_os_b4], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[72:75], %[v_os_b4], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[76:79], %[v_os_b4], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[80:83], %[v_os_b5], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[84:87], %[v_os_b5], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[88:91], %[v_os_b5], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[92:95], %[v_os_b5], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[96:99], %[v_os_b6], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[100:103], %[v_os_b6], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[104:107], %[v_os_b6], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[108:111], %[v_os_b6], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[112:115], %[v_os_b7], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[116:119], %[v_os_b7], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[120:123], %[v_os_b7], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[124:127], %[v_os_b7], s[12:15], 0 offen offset:3072 \n"
" s_cmp_gt_i32 %[s_loop_cnt] 1 ; move b with cond \n"
" s_cselect_b32 s86, %[s_tile_os_b], 0 \n"
" s_add_u32 s12, s86, s12 \n"
" s_addc_u32 s13, 0, s13 \n"
" s_waitcnt 0 \n"
"L_start%=: \n"
" s_waitcnt vmcnt(32) \n"
" s_barrier \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[0:1], v[128:129], 0 \n"
" buffer_load_dwordx4 acc[128:131], %[v_os_b0], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[2:3], v[130:131], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[4:5], v[132:133], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[6:7], v[134:135], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[8:9], v[136:137], [%[c0], %[c1], %[c2], %[c3]] \n"
" buffer_load_dwordx4 acc[132:135], %[v_os_b0], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[10:11], v[138:139], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[12:13], v[140:141], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[14:15], v[142:143], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[0:1], v[192:193], 0 \n"
" buffer_load_dwordx4 acc[136:139], %[v_os_b0], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[2:3], v[194:195], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[4:5], v[196:197], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[6:7], v[198:199], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[8:9], v[200:201], [%[c4], %[c5], %[c6], %[c7]] \n"
" buffer_load_dwordx4 acc[140:143], %[v_os_b0], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[10:11], v[202:203], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[12:13], v[204:205], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[14:15], v[206:207], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[16:17], v[128:129], 0 \n"
" buffer_load_dwordx4 acc[144:147], %[v_os_b1], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[18:19], v[130:131], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[20:21], v[132:133], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[22:23], v[134:135], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[24:25], v[136:137], [%[c8], %[c9], %[c10], %[c11]] \n"
" buffer_load_dwordx4 acc[148:151], %[v_os_b1], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[26:27], v[138:139], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[28:29], v[140:141], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[30:31], v[142:143], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[16:17], v[192:193], 0 \n"
" buffer_load_dwordx4 acc[152:155], %[v_os_b1], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[18:19], v[194:195], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[20:21], v[196:197], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[22:23], v[198:199], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[24:25], v[200:201], [%[c12], %[c13], %[c14], %[c15]] \n"
" buffer_load_dwordx4 acc[156:159], %[v_os_b1], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[26:27], v[202:203], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[28:29], v[204:205], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[30:31], v[206:207], [%[c12], %[c13], %[c14], %[c15]] \n"
" s_waitcnt vmcnt(32) \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[32:33], v[144:145], [%[c0], %[c1], %[c2], %[c3]] \n"
" buffer_load_dwordx4 acc[160:163], %[v_os_b2], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[34:35], v[146:147], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[36:37], v[148:149], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[38:39], v[150:151], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[40:41], v[152:153], [%[c0], %[c1], %[c2], %[c3]] \n"
" buffer_load_dwordx4 acc[164:167], %[v_os_b2], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[42:43], v[154:155], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[44:45], v[156:157], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[46:47], v[158:159], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[32:33], v[208:209], [%[c4], %[c5], %[c6], %[c7]] \n"
" buffer_load_dwordx4 acc[168:171], %[v_os_b2], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[34:35], v[210:211], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[36:37], v[212:213], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[38:39], v[214:215], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[40:41], v[216:217], [%[c4], %[c5], %[c6], %[c7]] \n"
" buffer_load_dwordx4 acc[172:175], %[v_os_b2], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[42:43], v[218:219], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[44:45], v[220:221], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[46:47], v[222:223], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[48:49], v[144:145], [%[c8], %[c9], %[c10], %[c11]] \n"
" buffer_load_dwordx4 acc[176:179], %[v_os_b3], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[50:51], v[146:147], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[52:53], v[148:149], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[54:55], v[150:151], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[56:57], v[152:153], [%[c8], %[c9], %[c10], %[c11]] \n"
" buffer_load_dwordx4 acc[180:183], %[v_os_b3], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[58:59], v[154:155], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[60:61], v[156:157], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[62:63], v[158:159], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[48:49], v[208:209], [%[c12], %[c13], %[c14], %[c15]] \n"
" buffer_load_dwordx4 acc[184:187], %[v_os_b3], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[50:51], v[210:211], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[52:53], v[212:213], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[54:55], v[214:215], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[56:57], v[216:217], [%[c12], %[c13], %[c14], %[c15]] \n"
" buffer_load_dwordx4 acc[188:191], %[v_os_b3], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[58:59], v[218:219], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[60:61], v[220:221], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[62:63], v[222:223], [%[c12], %[c13], %[c14], %[c15]] \n"
" s_waitcnt vmcnt(32) \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[64:65], v[160:161], [%[c0], %[c1], %[c2], %[c3]] \n"
" buffer_load_dwordx4 acc[192:195], %[v_os_b4], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[66:67], v[162:163], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[68:69], v[164:165], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[70:71], v[166:167], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[72:73], v[168:169], [%[c0], %[c1], %[c2], %[c3]] \n"
" buffer_load_dwordx4 acc[196:199], %[v_os_b4], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[74:75], v[170:171], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[76:77], v[172:173], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[78:79], v[174:175], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[64:65], v[224:225], [%[c4], %[c5], %[c6], %[c7]] \n"
" buffer_load_dwordx4 acc[200:203], %[v_os_b4], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[66:67], v[226:227], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[68:69], v[228:229], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[70:71], v[230:231], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[72:73], v[232:233], [%[c4], %[c5], %[c6], %[c7]] \n"
" buffer_load_dwordx4 acc[204:207], %[v_os_b4], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[74:75], v[234:235], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[76:77], v[236:237], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[78:79], v[238:239], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[80:81], v[160:161], [%[c8], %[c9], %[c10], %[c11]] \n"
" buffer_load_dwordx4 acc[208:211], %[v_os_b5], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[82:83], v[162:163], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[84:85], v[164:165], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[86:87], v[166:167], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[88:89], v[168:169], [%[c8], %[c9], %[c10], %[c11]] \n"
" buffer_load_dwordx4 acc[212:215], %[v_os_b5], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[90:91], v[170:171], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[92:93], v[172:173], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[94:95], v[174:175], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[80:81], v[224:225], [%[c12], %[c13], %[c14], %[c15]] \n"
" buffer_load_dwordx4 acc[216:219], %[v_os_b5], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[82:83], v[226:227], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[84:85], v[228:229], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[86:87], v[230:231], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[88:89], v[232:233], [%[c12], %[c13], %[c14], %[c15]] \n"
" buffer_load_dwordx4 acc[220:223], %[v_os_b5], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[90:91], v[234:235], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[92:93], v[236:237], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[94:95], v[238:239], [%[c12], %[c13], %[c14], %[c15]] \n"
" s_waitcnt vmcnt(32) \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[96:97], v[176:177], [%[c0], %[c1], %[c2], %[c3]] \n"
" buffer_load_dwordx4 acc[224:227], %[v_os_b6], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[98:99], v[178:179], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[100:101], v[180:181], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[102:103], v[182:183], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[104:105], v[184:185], [%[c0], %[c1], %[c2], %[c3]] \n"
" buffer_load_dwordx4 acc[228:231], %[v_os_b6], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[106:107], v[186:187], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[108:109], v[188:189], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[110:111], v[190:191], [%[c0], %[c1], %[c2], %[c3]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[96:97], v[240:241], [%[c4], %[c5], %[c6], %[c7]] \n"
" buffer_load_dwordx4 acc[232:235], %[v_os_b6], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[98:99], v[242:243], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[100:101], v[244:245], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[102:103], v[246:247], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[104:105], v[248:249], [%[c4], %[c5], %[c6], %[c7]] \n"
" buffer_load_dwordx4 acc[236:239], %[v_os_b6], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[106:107], v[250:251], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[108:109], v[252:253], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[110:111], v[254:255], [%[c4], %[c5], %[c6], %[c7]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[112:113], v[176:177], [%[c8], %[c9], %[c10], %[c11]] \n"
" buffer_load_dwordx4 acc[240:243], %[v_os_b7], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[114:115], v[178:179], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[116:117], v[180:181], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[118:119], v[182:183], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[120:121], v[184:185], [%[c8], %[c9], %[c10], %[c11]] \n"
" buffer_load_dwordx4 acc[244:247], %[v_os_b7], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[122:123], v[186:187], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[124:125], v[188:189], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[126:127], v[190:191], [%[c8], %[c9], %[c10], %[c11]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[112:113], v[240:241], [%[c12], %[c13], %[c14], %[c15]] \n"
" buffer_load_dwordx4 acc[248:251], %[v_os_b7], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[114:115], v[242:243], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[116:117], v[244:245], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[118:119], v[246:247], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[120:121], v[248:249], [%[c12], %[c13], %[c14], %[c15]] \n"
" buffer_load_dwordx4 acc[252:255], %[v_os_b7], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[122:123], v[250:251], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[124:125], v[252:253], [%[c12], %[c13], %[c14], %[c15]] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[126:127], v[254:255], [%[c12], %[c13], %[c14], %[c15]]\n"
" v_mul_f32 %[c0], %[scale_0], %[c0] \n"
" v_mul_f32 %[c1], %[scale_0], %[c1] \n"
" v_mul_f32 %[c2], %[scale_0], %[c2] \n"
" v_mul_f32 %[c3], %[scale_0], %[c3] \n"
" v_mul_f32 %[c4], %[scale_1], %[c4] \n"
" v_mul_f32 %[c5], %[scale_1], %[c5] \n"
" v_mul_f32 %[c6], %[scale_1], %[c6] \n"
" v_mul_f32 %[c7], %[scale_1], %[c7] \n"
" v_mul_f32 %[c8], %[scale_0], %[c8] \n"
" v_mul_f32 %[c9], %[scale_0], %[c9] \n"
" v_mul_f32 %[c10], %[scale_0], %[c10] \n"
" v_mul_f32 %[c11], %[scale_0], %[c11] \n"
" v_mul_f32 %[c12], %[scale_1], %[c12] \n"
" v_mul_f32 %[c13], %[scale_1], %[c13] \n"
" v_mul_f32 %[c14], %[scale_1], %[c14] \n"
" v_mul_f32 %[c15], %[scale_1], %[c15] \n"
_UK_PK_CVT_("%[c0]", "%[c1]", "%[c0]")
_UK_PK_CVT_("%[c2]", "%[c3]", "%[c1]")
_UK_PK_CVT_("%[c4]", "%[c5]", "%[c2]")
_UK_PK_CVT_("%[c6]", "%[c7]", "%[c3]")
_UK_PK_CVT_("%[c8]", "%[c9]", "%[c4]")
_UK_PK_CVT_("%[c10]", "%[c11]", "%[c5]")
_UK_PK_CVT_("%[c12]", "%[c13]", "%[c6]")
_UK_PK_CVT_("%[c14]", "%[c15]", "%[c7]")
" ;------------------------------ \n"
" ds_write_b64 %[v_sfl_sst], [%[c0],%[c1]] offset:0 + %[shfl_base] \n"
" ds_write_b64 %[v_sfl_sst], [%[c2],%[c3]] offset:4352 + %[shfl_base] \n"
" ds_write_b64 %[v_sfl_sst], [%[c4],%[c5]] offset:2176 + %[shfl_base] \n"
" ds_write_b64 %[v_sfl_sst], [%[c6],%[c7]] offset:6528 + %[shfl_base] \n"
" s_waitcnt lgkmcnt(0) \n"
" s_barrier \n"
" ds_read_b32 %[c0], %[v_sfl_sld] offset:0 + %[shfl_base] \n"
" ds_read_b32 %[c1], %[v_sfl_sld] offset:32 + %[shfl_base] \n"
" ds_read_b32 %[c2], %[v_sfl_sld] offset:64 + %[shfl_base] \n"
" ds_read_b32 %[c3], %[v_sfl_sld] offset:96 + %[shfl_base] \n"
" ds_read_b32 %[c4], %[v_sfl_sld] offset:4352 + %[shfl_base] \n"
" ds_read_b32 %[c5], %[v_sfl_sld] offset:4384 + %[shfl_base] \n"
" ds_read_b32 %[c6], %[v_sfl_sld] offset:4416 + %[shfl_base] \n"
" ds_read_b32 %[c7], %[v_sfl_sld] offset:4448 + %[shfl_base] \n"
" s_waitcnt lgkmcnt(0) \n"
" s_mov_b64 exec, %[s_execflag_0] \n"
_UK_ATOMIC_ADD_ " %[v_os_o0], %[c0], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_1] \n"
_UK_ATOMIC_ADD_ " %[v_os_o1], %[c1], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_2] \n"
_UK_ATOMIC_ADD_ " %[v_os_o2], %[c2], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_3] \n"
_UK_ATOMIC_ADD_ " %[v_os_o3], %[c3], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_4] \n"
_UK_ATOMIC_ADD_ " %[v_os_o4], %[c4], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_5] \n"
_UK_ATOMIC_ADD_ " %[v_os_o5], %[c5], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_6] \n"
_UK_ATOMIC_ADD_ " %[v_os_o6], %[c6], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_7] \n"
_UK_ATOMIC_ADD_ " %[v_os_o7], %[c7], s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_sub_i32 %[s_loop_cnt], %[s_loop_cnt], 1 ; k-- \n"
" s_cmp_gt_i32 %[s_loop_cnt] 0 \n"
" s_cbranch_scc0 L_end%= \n"
" s_cmp_gt_i32 %[s_loop_cnt] 1 ; move b with cond \n"
" s_cselect_b32 s86, %[s_tile_os_b], 0 \n"
" s_add_u32 s12, s86, s12 \n"
" s_addc_u32 s13, 0, s13 \n"
" s_add_u32 s8, %[s_tile_os_o], s8 \n"
" s_addc_u32 s9, 0, s9 \n"
" s_waitcnt vmcnt(32) \n"
" s_barrier \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[128:129], v[128:129], 0 \n"
" buffer_load_dwordx4 acc[0:3], %[v_os_b0], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[130:131], v[130:131], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[132:133], v[132:133], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[134:135], v[134:135], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[136:137], v[136:137], [%[c16],%[c17],%[c18],%[c19]] \n"
" buffer_load_dwordx4 acc[4:7], %[v_os_b0], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[138:139], v[138:139], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[140:141], v[140:141], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[142:143], v[142:143], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[128:129], v[192:193], 0 \n"
" buffer_load_dwordx4 acc[8:11], %[v_os_b0], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[130:131], v[194:195], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[132:133], v[196:197], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[134:135], v[198:199], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[136:137], v[200:201], [%[c20],%[c21],%[c22],%[c23]] \n"
" buffer_load_dwordx4 acc[12:15], %[v_os_b0], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[138:139], v[202:203], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[140:141], v[204:205], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[142:143], v[206:207], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[144:145], v[128:129], 0 \n"
" buffer_load_dwordx4 acc[16:19], %[v_os_b1], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[146:147], v[130:131], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[148:149], v[132:133], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[150:151], v[134:135], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[152:153], v[136:137], [%[c24],%[c25],%[c26],%[c27]] \n"
" buffer_load_dwordx4 acc[20:23], %[v_os_b1], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[154:155], v[138:139], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[156:157], v[140:141], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[158:159], v[142:143], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[144:145], v[192:193], 0 \n"
" buffer_load_dwordx4 acc[24:27], %[v_os_b1], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[146:147], v[194:195], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[148:149], v[196:197], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[150:151], v[198:199], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[152:153], v[200:201], [%[c28],%[c29],%[c30],%[c31]] \n"
" buffer_load_dwordx4 acc[28:31], %[v_os_b1], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[154:155], v[202:203], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[156:157], v[204:205], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[158:159], v[206:207], [%[c28],%[c29],%[c30],%[c31]] \n"
" s_waitcnt vmcnt(32) \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[160:161], v[144:145], [%[c16],%[c17],%[c18],%[c19]] \n"
" buffer_load_dwordx4 acc[32:35], %[v_os_b2], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[162:163], v[146:147], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[164:165], v[148:149], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[166:167], v[150:151], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[168:169], v[152:153], [%[c16],%[c17],%[c18],%[c19]] \n"
" buffer_load_dwordx4 acc[36:39], %[v_os_b2], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[170:171], v[154:155], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[172:173], v[156:157], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[174:175], v[158:159], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[160:161], v[208:209], [%[c20],%[c21],%[c22],%[c23]] \n"
" buffer_load_dwordx4 acc[40:43], %[v_os_b2], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[162:163], v[210:211], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[164:165], v[212:213], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[166:167], v[214:215], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[168:169], v[216:217], [%[c20],%[c21],%[c22],%[c23]] \n"
" buffer_load_dwordx4 acc[44:47], %[v_os_b2], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[170:171], v[218:219], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[172:173], v[220:221], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[174:175], v[222:223], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[176:177], v[144:145], [%[c24],%[c25],%[c26],%[c27]] \n"
" buffer_load_dwordx4 acc[48:51], %[v_os_b3], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[178:179], v[146:147], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[180:181], v[148:149], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[182:183], v[150:151], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[184:185], v[152:153], [%[c24],%[c25],%[c26],%[c27]] \n"
" buffer_load_dwordx4 acc[52:55], %[v_os_b3], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[186:187], v[154:155], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[188:189], v[156:157], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[190:191], v[158:159], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[176:177], v[208:209], [%[c28],%[c29],%[c30],%[c31]] \n"
" buffer_load_dwordx4 acc[56:59], %[v_os_b3], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[178:179], v[210:211], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[180:181], v[212:213], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[182:183], v[214:215], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[184:185], v[216:217], [%[c28],%[c29],%[c30],%[c31]] \n"
" buffer_load_dwordx4 acc[60:63], %[v_os_b3], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[186:187], v[218:219], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[188:189], v[220:221], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[190:191], v[222:223], [%[c28],%[c29],%[c30],%[c31]] \n"
" s_waitcnt vmcnt(32) \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[192:193], v[160:161], [%[c16],%[c17],%[c18],%[c19]] \n"
" buffer_load_dwordx4 acc[64:67], %[v_os_b4], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[194:195], v[162:163], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[196:197], v[164:165], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[198:199], v[166:167], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[200:201], v[168:169], [%[c16],%[c17],%[c18],%[c19]] \n"
" buffer_load_dwordx4 acc[68:71], %[v_os_b4], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[202:203], v[170:171], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[204:205], v[172:173], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[206:207], v[174:175], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[192:193], v[224:225], [%[c20],%[c21],%[c22],%[c23]] \n"
" buffer_load_dwordx4 acc[72:75], %[v_os_b4], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[194:195], v[226:227], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[196:197], v[228:229], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[198:199], v[230:231], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[200:201], v[232:233], [%[c20],%[c21],%[c22],%[c23]] \n"
" buffer_load_dwordx4 acc[76:79], %[v_os_b4], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[202:203], v[234:235], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[204:205], v[236:237], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[206:207], v[238:239], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[208:209], v[160:161], [%[c24],%[c25],%[c26],%[c27]] \n"
" buffer_load_dwordx4 acc[80:83], %[v_os_b5], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[210:211], v[162:163], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[212:213], v[164:165], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[214:215], v[166:167], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[216:217], v[168:169], [%[c24],%[c25],%[c26],%[c27]] \n"
" buffer_load_dwordx4 acc[84:87], %[v_os_b5], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[218:219], v[170:171], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[220:221], v[172:173], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[222:223], v[174:175], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[208:209], v[224:225], [%[c28],%[c29],%[c30],%[c31]] \n"
" buffer_load_dwordx4 acc[88:91], %[v_os_b5], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[210:211], v[226:227], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[212:213], v[228:229], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[214:215], v[230:231], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[216:217], v[232:233], [%[c28],%[c29],%[c30],%[c31]] \n"
" buffer_load_dwordx4 acc[92:95], %[v_os_b5], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[218:219], v[234:235], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[220:221], v[236:237], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[222:223], v[238:239], [%[c28],%[c29],%[c30],%[c31]] \n"
" s_waitcnt vmcnt(32) \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[224:225], v[176:177], [%[c16],%[c17],%[c18],%[c19]] \n"
" buffer_load_dwordx4 acc[96:99], %[v_os_b6], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[226:227], v[178:179], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[228:229], v[180:181], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[230:231], v[182:183], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[232:233], v[184:185], [%[c16],%[c17],%[c18],%[c19]] \n"
" buffer_load_dwordx4 acc[100:103], %[v_os_b6], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[234:235], v[186:187], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[236:237], v[188:189], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[238:239], v[190:191], [%[c16],%[c17],%[c18],%[c19]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[224:225], v[240:241], [%[c20],%[c21],%[c22],%[c23]] \n"
" buffer_load_dwordx4 acc[104:107], %[v_os_b6], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[226:227], v[242:243], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[228:229], v[244:245], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[230:231], v[246:247], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[232:233], v[248:249], [%[c20],%[c21],%[c22],%[c23]] \n"
" buffer_load_dwordx4 acc[108:111], %[v_os_b6], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[234:235], v[250:251], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[236:237], v[252:253], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[238:239], v[254:255], [%[c20],%[c21],%[c22],%[c23]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[240:241], v[176:177], [%[c24],%[c25],%[c26],%[c27]] \n"
" buffer_load_dwordx4 acc[112:115], %[v_os_b7], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[242:243], v[178:179], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[244:245], v[180:181], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[246:247], v[182:183], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[248:249], v[184:185], [%[c24],%[c25],%[c26],%[c27]] \n"
" buffer_load_dwordx4 acc[116:119], %[v_os_b7], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[250:251], v[186:187], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[252:253], v[188:189], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[254:255], v[190:191], [%[c24],%[c25],%[c26],%[c27]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[240:241], v[240:241], [%[c28],%[c29],%[c30],%[c31]] \n"
" buffer_load_dwordx4 acc[120:123], %[v_os_b7], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[242:243], v[242:243], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[244:245], v[244:245], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[246:247], v[246:247], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[248:249], v[248:249], [%[c28],%[c29],%[c30],%[c31]] \n"
" buffer_load_dwordx4 acc[124:127], %[v_os_b7], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[250:251], v[250:251], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[252:253], v[252:253], [%[c28],%[c29],%[c30],%[c31]] \n"
_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[254:255], v[254:255], [%[c28],%[c29],%[c30],%[c31]]\n"
" v_mul_f32 %[c16], %[scale_0], %[c16] \n"
" v_mul_f32 %[c17], %[scale_0], %[c17] \n"
" v_mul_f32 %[c18], %[scale_0], %[c18] \n"
" v_mul_f32 %[c19], %[scale_0], %[c19] \n"
" v_mul_f32 %[c20], %[scale_1], %[c20] \n"
" v_mul_f32 %[c21], %[scale_1], %[c21] \n"
" v_mul_f32 %[c22], %[scale_1], %[c22] \n"
" v_mul_f32 %[c23], %[scale_1], %[c23] \n"
" v_mul_f32 %[c24], %[scale_0], %[c24] \n"
" v_mul_f32 %[c25], %[scale_0], %[c25] \n"
" v_mul_f32 %[c26], %[scale_0], %[c26] \n"
" v_mul_f32 %[c27], %[scale_0], %[c27] \n"
" v_mul_f32 %[c28], %[scale_1], %[c28] \n"
" v_mul_f32 %[c29], %[scale_1], %[c29] \n"
" v_mul_f32 %[c30], %[scale_1], %[c30] \n"
" v_mul_f32 %[c31], %[scale_1], %[c31] \n"
" s_mov_b32 s52, 0x07060302 ; v_perm\n"
" s_mov_b64 s[38:39], exec ; save current exec\n"
" s_mov_b32 s8, %[s_res_o0] \n"
" s_mov_b32 s9, %[s_res_o1] \n"
" s_mov_b32 s12, %[s_res_b0] \n"
" s_mov_b32 s13, %[s_res_b1] \n"
" s_mov_b32 s14, %[s_res_b2] \n"
" s_mov_b32 s15, %[s_res_b3] \n"
" ds_read_b64 v[128:129], %[v_sld_y_os] offset:0 + %[sld_a_base] \n"
" ds_read_b64 v[130:131], %[v_sld_y_os] offset:128 + %[sld_a_base] \n"
" ds_read_b64 v[132:133], %[v_sld_y_os] offset:1024 + %[sld_a_base] \n"
" ds_read_b64 v[134:135], %[v_sld_y_os] offset:1152 + %[sld_a_base] \n"
" ds_read_b64 v[136:137], %[v_sld_y_os] offset:2048 + %[sld_a_base] \n"
" ds_read_b64 v[138:139], %[v_sld_y_os] offset:2176 + %[sld_a_base] \n"
" ds_read_b64 v[140:141], %[v_sld_y_os] offset:3072 + %[sld_a_base] \n"
" ds_read_b64 v[142:143], %[v_sld_y_os] offset:3200 + %[sld_a_base] \n"
" ds_read_b64 v[144:145], %[v_sld_y_os] offset:4096 + %[sld_a_base] \n"
" ds_read_b64 v[146:147], %[v_sld_y_os] offset:4224 + %[sld_a_base] \n"
" ds_read_b64 v[148:149], %[v_sld_y_os] offset:5120 + %[sld_a_base] \n"
" ds_read_b64 v[150:151], %[v_sld_y_os] offset:5248 + %[sld_a_base] \n"
" ds_read_b64 v[152:153], %[v_sld_y_os] offset:6144 + %[sld_a_base] \n"
" ds_read_b64 v[154:155], %[v_sld_y_os] offset:6272 + %[sld_a_base] \n"
" ds_read_b64 v[156:157], %[v_sld_y_os] offset:7168 + %[sld_a_base] \n"
" ds_read_b64 v[158:159], %[v_sld_y_os] offset:7296 + %[sld_a_base] \n"
" ds_read_b64 v[160:161], %[v_sld_y_os] offset:8192 + %[sld_a_base] \n"
" ds_read_b64 v[162:163], %[v_sld_y_os] offset:8320 + %[sld_a_base] \n"
" ds_read_b64 v[164:165], %[v_sld_y_os] offset:9216 + %[sld_a_base] \n"
" ds_read_b64 v[166:167], %[v_sld_y_os] offset:9344 + %[sld_a_base] \n"
" ds_read_b64 v[168:169], %[v_sld_y_os] offset:10240 + %[sld_a_base] \n"
" ds_read_b64 v[170:171], %[v_sld_y_os] offset:10368 + %[sld_a_base] \n"
" ds_read_b64 v[172:173], %[v_sld_y_os] offset:11264 + %[sld_a_base] \n"
" ds_read_b64 v[174:175], %[v_sld_y_os] offset:11392 + %[sld_a_base] \n"
" ds_read_b64 v[176:177], %[v_sld_y_os] offset:12288 + %[sld_a_base] \n"
" ds_read_b64 v[178:179], %[v_sld_y_os] offset:12416 + %[sld_a_base] \n"
" ds_read_b64 v[180:181], %[v_sld_y_os] offset:13312 + %[sld_a_base] \n"
" ds_read_b64 v[182:183], %[v_sld_y_os] offset:13440 + %[sld_a_base] \n"
" ds_read_b64 v[184:185], %[v_sld_y_os] offset:14336 + %[sld_a_base] \n"
" ds_read_b64 v[186:187], %[v_sld_y_os] offset:14464 + %[sld_a_base] \n"
" ds_read_b64 v[188:189], %[v_sld_y_os] offset:15360 + %[sld_a_base] \n"
" ds_read_b64 v[190:191], %[v_sld_y_os] offset:15488 + %[sld_a_base] \n"
" ds_read_b64 v[192:193], %[v_sld_y_os] offset:16384 + %[sld_a_base] \n"
" ds_read_b64 v[194:195], %[v_sld_y_os] offset:16512 + %[sld_a_base] \n"
" ds_read_b64 v[196:197], %[v_sld_y_os] offset:17408 + %[sld_a_base] \n"
" ds_read_b64 v[198:199], %[v_sld_y_os] offset:17536 + %[sld_a_base] \n"
" ds_read_b64 v[200:201], %[v_sld_y_os] offset:18432 + %[sld_a_base] \n"
" ds_read_b64 v[202:203], %[v_sld_y_os] offset:18560 + %[sld_a_base] \n"
" ds_read_b64 v[204:205], %[v_sld_y_os] offset:19456 + %[sld_a_base] \n"
" ds_read_b64 v[206:207], %[v_sld_y_os] offset:19584 + %[sld_a_base] \n"
" ds_read_b64 v[208:209], %[v_sld_y_os] offset:20480 + %[sld_a_base] \n"
" ds_read_b64 v[210:211], %[v_sld_y_os] offset:20608 + %[sld_a_base] \n"
" ds_read_b64 v[212:213], %[v_sld_y_os] offset:21504 + %[sld_a_base] \n"
" ds_read_b64 v[214:215], %[v_sld_y_os] offset:21632 + %[sld_a_base] \n"
" ds_read_b64 v[216:217], %[v_sld_y_os] offset:22528 + %[sld_a_base] \n"
" ds_read_b64 v[218:219], %[v_sld_y_os] offset:22656 + %[sld_a_base] \n"
" ds_read_b64 v[220:221], %[v_sld_y_os] offset:23552 + %[sld_a_base] \n"
" ds_read_b64 v[222:223], %[v_sld_y_os] offset:23680 + %[sld_a_base] \n"
" ds_read_b64 v[224:225], %[v_sld_y_os] offset:24576 + %[sld_a_base] \n"
" ds_read_b64 v[226:227], %[v_sld_y_os] offset:24704 + %[sld_a_base] \n"
" ds_read_b64 v[228:229], %[v_sld_y_os] offset:25600 + %[sld_a_base] \n"
" ds_read_b64 v[230:231], %[v_sld_y_os] offset:25728 + %[sld_a_base] \n"
" ds_read_b64 v[232:233], %[v_sld_y_os] offset:26624 + %[sld_a_base] \n"
" ds_read_b64 v[234:235], %[v_sld_y_os] offset:26752 + %[sld_a_base] \n"
" ds_read_b64 v[236:237], %[v_sld_y_os] offset:27648 + %[sld_a_base] \n"
" ds_read_b64 v[238:239], %[v_sld_y_os] offset:27776 + %[sld_a_base] \n"
" ds_read_b64 v[240:241], %[v_sld_y_os] offset:28672 + %[sld_a_base] \n"
" ds_read_b64 v[242:243], %[v_sld_y_os] offset:28800 + %[sld_a_base] \n"
" ds_read_b64 v[244:245], %[v_sld_y_os] offset:29696 + %[sld_a_base] \n"
" ds_read_b64 v[246:247], %[v_sld_y_os] offset:29824 + %[sld_a_base] \n"
" ds_read_b64 v[248:249], %[v_sld_y_os] offset:30720 + %[sld_a_base] \n"
" ds_read_b64 v[250:251], %[v_sld_y_os] offset:30848 + %[sld_a_base] \n"
" ds_read_b64 v[252:253], %[v_sld_y_os] offset:31744 + %[sld_a_base] \n"
" ds_read_b64 v[254:255], %[v_sld_y_os] offset:31872 + %[sld_a_base] \n"
" s_waitcnt 0 \n"
" buffer_load_dwordx4 acc[0:3], %[v_os_b0], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[4:7], %[v_os_b0], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[8:11], %[v_os_b0], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[12:15], %[v_os_b0], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[16:19], %[v_os_b1], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[20:23], %[v_os_b1], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[24:27], %[v_os_b1], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[28:31], %[v_os_b1], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[32:35], %[v_os_b2], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[36:39], %[v_os_b2], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[40:43], %[v_os_b2], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[44:47], %[v_os_b2], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[48:51], %[v_os_b3], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[52:55], %[v_os_b3], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[56:59], %[v_os_b3], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[60:63], %[v_os_b3], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[64:67], %[v_os_b4], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[68:71], %[v_os_b4], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[72:75], %[v_os_b4], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[76:79], %[v_os_b4], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[80:83], %[v_os_b5], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[84:87], %[v_os_b5], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[88:91], %[v_os_b5], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[92:95], %[v_os_b5], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[96:99], %[v_os_b6], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[100:103], %[v_os_b6], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[104:107], %[v_os_b6], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[108:111], %[v_os_b6], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[112:115], %[v_os_b7], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[116:119], %[v_os_b7], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[120:123], %[v_os_b7], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[124:127], %[v_os_b7], s[12:15], 0 offen offset:3072 \n"
" s_cmp_gt_i32 %[s_loop_cnt] 1 ; move b with cond \n"
" s_cselect_b32 s86, %[s_tile_os_b], 0 \n"
" s_add_u32 s12, s86, s12 \n"
" s_addc_u32 s13, 0, s13 \n"
" s_waitcnt 0 \n"
"L_start%=: \n"
" s_waitcnt vmcnt(32) \n"
" s_barrier \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[0:1], v[128:129], 0 \n"
" buffer_load_dwordx4 acc[128:131], %[v_os_b0], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[2:3], v[130:131], [%[c0], %[c1], %[c2], %[c3]] "
"\n" _UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[4:5], v[132:133], [%[c0], %[c1], %[c2], "
"%[c3]] \n" _UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[6:7], v[134:135], "
"[%[c0], %[c1], %[c2], %[c3]] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[8:9], v[136:137], [%[c0], %[c1], %[c2], %[c3]] \n"
" buffer_load_dwordx4 acc[132:135], %[v_os_b0], s[12:15], 0 offen offset:1024 \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[10:11], v[138:139], [%[c0], %[c1], %[c2], %[c3]] "
"\n" _UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[12:13], v[140:141], [%[c0], %[c1], %[c2], "
"%[c3]] \n" _UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[14:15], v[142:143], "
"[%[c0], %[c1], %[c2], %[c3]] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[0:1], v[192:193], 0 \n"
" buffer_load_dwordx4 acc[136:139], %[v_os_b0], s[12:15], 0 offen offset:2048 \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[2:3], v[194:195], [%[c4], %[c5], %[c6], %[c7]] "
"\n" _UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[4:5], v[196:197], [%[c4], %[c5], %[c6], "
"%[c7]] \n" _UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[6:7], v[198:199], "
"[%[c4], %[c5], %[c6], %[c7]] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[8:9], v[200:201], [%[c4], %[c5], %[c6], %[c7]] \n"
" buffer_load_dwordx4 acc[140:143], %[v_os_b0], s[12:15], 0 offen offset:3072 \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[10:11], v[202:203], [%[c4], %[c5], %[c6], %[c7]] "
"\n" _UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[12:13], v[204:205], [%[c4], %[c5], %[c6], "
"%[c7]] \n" _UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[14:15], v[206:207], "
"[%[c4], %[c5], %[c6], %[c7]] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[16:17], v[128:129], 0 \n"
" buffer_load_dwordx4 acc[144:147], %[v_os_b1], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[18:19], v[130:131], [%[c8], %[c9], %[c10], %[c11]] "
"\n" _UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[20:21], v[132:133], [%[c8], %[c9], "
"%[c10], %[c11]] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[22:23], v[134:135], [%[c8], %[c9], %[c10], %[c11]] "
"\n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[24:25], v[136:137], [%[c8], %[c9], %[c10], %[c11]] \n"
" buffer_load_dwordx4 acc[148:151], %[v_os_b1], s[12:15], 0 offen offset:1024 \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[26:27], v[138:139], [%[c8], %[c9], %[c10], %[c11]] "
"\n" _UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[28:29], v[140:141], [%[c8], %[c9], "
"%[c10], %[c11]] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[30:31], v[142:143], [%[c8], %[c9], %[c10], %[c11]] "
"\n" _UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[16:17], v[192:193], 0 \n"
" buffer_load_dwordx4 acc[152:155], %[v_os_b1], s[12:15], 0 offen offset:2048 \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[18:19], v[194:195], [%[c12], %[c13], %[c14], %[c15]] "
"\n" _UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[20:21], v[196:197], [%[c12], %[c13], "
"%[c14], %[c15]] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[22:23], v[198:199], [%[c12], %[c13], %[c14], %[c15]] "
"\n" _UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[24:25], v[200:201], [%[c12], %[c13], "
"%[c14], %[c15]] \n"
" buffer_load_dwordx4 acc[156:159], %[v_os_b1], s[12:15], 0 offen offset:3072 \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[26:27], v[202:203], [%[c12], %[c13], %[c14], %[c15]] "
"\n" _UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[28:29], v[204:205], [%[c12], %[c13], "
"%[c14], %[c15]] \n" _UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[30:31], "
"v[206:207], [%[c12], %[c13], %[c14], %[c15]] \n"
" s_waitcnt vmcnt(32) \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[32:33], v[144:145], [%[c0], %[c1], %[c2], %[c3]] \n"
" buffer_load_dwordx4 acc[160:163], %[v_os_b2], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[34:35], v[146:147], [%[c0], %[c1], %[c2], %[c3]] "
"\n" _UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[36:37], v[148:149], [%[c0], %[c1], %[c2], "
"%[c3]] \n" _UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[38:39], v[150:151], "
"[%[c0], %[c1], %[c2], %[c3]] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[40:41], v[152:153], [%[c0], %[c1], %[c2], %[c3]] \n"
" buffer_load_dwordx4 acc[164:167], %[v_os_b2], s[12:15], 0 offen offset:1024 \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[42:43], v[154:155], [%[c0], %[c1], %[c2], %[c3]] "
"\n" _UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[44:45], v[156:157], [%[c0], %[c1], %[c2], "
"%[c3]] \n" _UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[46:47], v[158:159], "
"[%[c0], %[c1], %[c2], %[c3]] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[32:33], v[208:209], [%[c4], %[c5], %[c6], %[c7]] \n"
" buffer_load_dwordx4 acc[168:171], %[v_os_b2], s[12:15], 0 offen offset:2048 \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[34:35], v[210:211], [%[c4], %[c5], %[c6], %[c7]] "
"\n" _UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[36:37], v[212:213], [%[c4], %[c5], %[c6], "
"%[c7]] \n" _UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[38:39], v[214:215], "
"[%[c4], %[c5], %[c6], %[c7]] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[40:41], v[216:217], [%[c4], %[c5], %[c6], %[c7]] \n"
" buffer_load_dwordx4 acc[172:175], %[v_os_b2], s[12:15], 0 offen offset:3072 \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[42:43], v[218:219], [%[c4], %[c5], %[c6], %[c7]] "
"\n" _UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[44:45], v[220:221], [%[c4], %[c5], %[c6], "
"%[c7]] \n" _UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[46:47], v[222:223], "
"[%[c4], %[c5], %[c6], %[c7]] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[48:49], v[144:145], [%[c8], %[c9], %[c10], %[c11]] \n"
" buffer_load_dwordx4 acc[176:179], %[v_os_b3], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[50:51], v[146:147], [%[c8], %[c9], %[c10], %[c11]] "
"\n" _UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[52:53], v[148:149], [%[c8], %[c9], "
"%[c10], %[c11]] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[54:55], v[150:151], [%[c8], %[c9], %[c10], %[c11]] "
"\n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[56:57], v[152:153], [%[c8], %[c9], %[c10], %[c11]] \n"
" buffer_load_dwordx4 acc[180:183], %[v_os_b3], s[12:15], 0 offen offset:1024 \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[58:59], v[154:155], [%[c8], %[c9], %[c10], %[c11]] "
"\n" _UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[60:61], v[156:157], [%[c8], %[c9], "
"%[c10], %[c11]] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[62:63], v[158:159], [%[c8], %[c9], %[c10], %[c11]] "
"\n" _UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[48:49], v[208:209], [%[c12], %[c13], "
"%[c14], %[c15]] \n"
" buffer_load_dwordx4 acc[184:187], %[v_os_b3], s[12:15], 0 offen offset:2048 \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[50:51], v[210:211], [%[c12], %[c13], %[c14], %[c15]] "
"\n" _UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[52:53], v[212:213], [%[c12], %[c13], "
"%[c14], %[c15]] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[54:55], v[214:215], [%[c12], %[c13], %[c14], %[c15]] "
"\n" _UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[56:57], v[216:217], [%[c12], %[c13], "
"%[c14], %[c15]] \n"
" buffer_load_dwordx4 acc[188:191], %[v_os_b3], s[12:15], 0 offen offset:3072 \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[58:59], v[218:219], [%[c12], %[c13], %[c14], %[c15]] "
"\n" _UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[60:61], v[220:221], [%[c12], %[c13], "
"%[c14], %[c15]] \n" _UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[62:63], "
"v[222:223], [%[c12], %[c13], %[c14], %[c15]] \n"
" s_waitcnt vmcnt(32) \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[64:65], v[160:161], [%[c0], %[c1], %[c2], %[c3]] \n"
" buffer_load_dwordx4 acc[192:195], %[v_os_b4], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[66:67], v[162:163], [%[c0], %[c1], %[c2], %[c3]] "
"\n" _UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[68:69], v[164:165], [%[c0], %[c1], %[c2], "
"%[c3]] \n" _UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[70:71], v[166:167], "
"[%[c0], %[c1], %[c2], %[c3]] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[72:73], v[168:169], [%[c0], %[c1], %[c2], %[c3]] \n"
" buffer_load_dwordx4 acc[196:199], %[v_os_b4], s[12:15], 0 offen offset:1024 \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[74:75], v[170:171], [%[c0], %[c1], %[c2], %[c3]] "
"\n" _UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[76:77], v[172:173], [%[c0], %[c1], %[c2], "
"%[c3]] \n" _UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[78:79], v[174:175], "
"[%[c0], %[c1], %[c2], %[c3]] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[64:65], v[224:225], [%[c4], %[c5], %[c6], %[c7]] \n"
" buffer_load_dwordx4 acc[200:203], %[v_os_b4], s[12:15], 0 offen offset:2048 \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[66:67], v[226:227], [%[c4], %[c5], %[c6], %[c7]] "
"\n" _UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[68:69], v[228:229], [%[c4], %[c5], %[c6], "
"%[c7]] \n" _UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[70:71], v[230:231], "
"[%[c4], %[c5], %[c6], %[c7]] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[72:73], v[232:233], [%[c4], %[c5], %[c6], %[c7]] \n"
" buffer_load_dwordx4 acc[204:207], %[v_os_b4], s[12:15], 0 offen offset:3072 \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[74:75], v[234:235], [%[c4], %[c5], %[c6], %[c7]] "
"\n" _UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[76:77], v[236:237], [%[c4], %[c5], %[c6], "
"%[c7]] \n" _UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[78:79], v[238:239], "
"[%[c4], %[c5], %[c6], %[c7]] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[80:81], v[160:161], [%[c8], %[c9], %[c10], %[c11]] \n"
" buffer_load_dwordx4 acc[208:211], %[v_os_b5], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[82:83], v[162:163], [%[c8], %[c9], %[c10], %[c11]] "
"\n" _UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[84:85], v[164:165], [%[c8], %[c9], "
"%[c10], %[c11]] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[86:87], v[166:167], [%[c8], %[c9], %[c10], %[c11]] "
"\n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[88:89], v[168:169], [%[c8], %[c9], %[c10], %[c11]] \n"
" buffer_load_dwordx4 acc[212:215], %[v_os_b5], s[12:15], 0 offen offset:1024 \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[90:91], v[170:171], [%[c8], %[c9], %[c10], %[c11]] "
"\n" _UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[92:93], v[172:173], [%[c8], %[c9], "
"%[c10], %[c11]] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[94:95], v[174:175], [%[c8], %[c9], %[c10], %[c11]] "
"\n" _UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[80:81], v[224:225], [%[c12], %[c13], "
"%[c14], %[c15]] \n"
" buffer_load_dwordx4 acc[216:219], %[v_os_b5], s[12:15], 0 offen offset:2048 \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[82:83], v[226:227], [%[c12], %[c13], %[c14], %[c15]] "
"\n" _UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[84:85], v[228:229], [%[c12], %[c13], "
"%[c14], %[c15]] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[86:87], v[230:231], [%[c12], %[c13], %[c14], %[c15]] "
"\n" _UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[88:89], v[232:233], [%[c12], %[c13], "
"%[c14], %[c15]] \n"
" buffer_load_dwordx4 acc[220:223], %[v_os_b5], s[12:15], 0 offen offset:3072 \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[90:91], v[234:235], [%[c12], %[c13], %[c14], %[c15]] "
"\n" _UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[92:93], v[236:237], [%[c12], %[c13], "
"%[c14], %[c15]] \n" _UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[94:95], "
"v[238:239], [%[c12], %[c13], %[c14], %[c15]] \n"
" s_waitcnt vmcnt(32) \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[96:97], v[176:177], [%[c0], %[c1], %[c2], %[c3]] \n"
" buffer_load_dwordx4 acc[224:227], %[v_os_b6], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[98:99], v[178:179], [%[c0], %[c1], %[c2], %[c3]] "
"\n" _UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[100:101], v[180:181], [%[c0], %[c1], "
"%[c2], %[c3]] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[102:103], v[182:183], [%[c0], %[c1], %[c2], %[c3]] "
"\n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[104:105], v[184:185], [%[c0], %[c1], %[c2], %[c3]] \n"
" buffer_load_dwordx4 acc[228:231], %[v_os_b6], s[12:15], 0 offen offset:1024 \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[106:107], v[186:187], [%[c0], %[c1], %[c2], %[c3]] "
"\n" _UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[108:109], v[188:189], [%[c0], %[c1], "
"%[c2], %[c3]] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[110:111], v[190:191], [%[c0], %[c1], %[c2], %[c3]] "
"\n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[96:97], v[240:241], [%[c4], %[c5], %[c6], %[c7]] \n"
" buffer_load_dwordx4 acc[232:235], %[v_os_b6], s[12:15], 0 offen offset:2048 \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[98:99], v[242:243], [%[c4], %[c5], %[c6], %[c7]] "
"\n" _UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[100:101], v[244:245], [%[c4], %[c5], "
"%[c6], %[c7]] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[102:103], v[246:247], [%[c4], %[c5], %[c6], %[c7]] "
"\n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[104:105], v[248:249], [%[c4], %[c5], %[c6], %[c7]] \n"
" buffer_load_dwordx4 acc[236:239], %[v_os_b6], s[12:15], 0 offen offset:3072 \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[106:107], v[250:251], [%[c4], %[c5], %[c6], %[c7]] "
"\n" _UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[108:109], v[252:253], [%[c4], %[c5], "
"%[c6], %[c7]] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[110:111], v[254:255], [%[c4], %[c5], %[c6], %[c7]] "
"\n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[112:113], v[176:177], [%[c8], %[c9], %[c10], %[c11]] \n"
" buffer_load_dwordx4 acc[240:243], %[v_os_b7], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[114:115], v[178:179], [%[c8], %[c9], %[c10], %[c11]] "
"\n" _UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[116:117], v[180:181], [%[c8], %[c9], "
"%[c10], %[c11]] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[118:119], v[182:183], [%[c8], %[c9], %[c10], %[c11]] "
"\n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[120:121], v[184:185], [%[c8], %[c9], %[c10], %[c11]] \n"
" buffer_load_dwordx4 acc[244:247], %[v_os_b7], s[12:15], 0 offen offset:1024 \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[122:123], v[186:187], [%[c8], %[c9], %[c10], %[c11]] "
"\n" _UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[124:125], v[188:189], [%[c8], %[c9], "
"%[c10], %[c11]] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[126:127], v[190:191], [%[c8], %[c9], %[c10], %[c11]] "
"\n" _UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[112:113], v[240:241], [%[c12], %[c13], "
"%[c14], %[c15]] \n"
" buffer_load_dwordx4 acc[248:251], %[v_os_b7], s[12:15], 0 offen offset:2048 \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[114:115], v[242:243], [%[c12], %[c13], %[c14], "
"%[c15]] \n" _UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[116:117], v[244:245], [%[c12], "
"%[c13], %[c14], %[c15]] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[118:119], v[246:247], [%[c12], %[c13], %[c14], "
"%[c15]] \n" _UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[120:121], v[248:249], [%[c12], "
"%[c13], %[c14], %[c15]] \n"
" buffer_load_dwordx4 acc[252:255], %[v_os_b7], s[12:15], 0 offen offset:3072 \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[122:123], v[250:251], [%[c12], %[c13], %[c14], "
"%[c15]] \n" _UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[124:125], v[252:253], [%[c12], "
"%[c13], %[c14], %[c15]] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[126:127], v[254:255], [%[c12], %[c13], %[c14], "
"%[c15]]\n"
" v_mul_f32 %[c0], %[scale_0], %[c0] \n"
" v_mul_f32 %[c1], %[scale_0], %[c1] \n"
" v_mul_f32 %[c2], %[scale_0], %[c2] \n"
" v_mul_f32 %[c3], %[scale_0], %[c3] \n"
" v_mul_f32 %[c4], %[scale_1], %[c4] \n"
" v_mul_f32 %[c5], %[scale_1], %[c5] \n"
" v_mul_f32 %[c6], %[scale_1], %[c6] \n"
" v_mul_f32 %[c7], %[scale_1], %[c7] \n"
" v_mul_f32 %[c8], %[scale_0], %[c8] \n"
" v_mul_f32 %[c9], %[scale_0], %[c9] \n"
" v_mul_f32 %[c10], %[scale_0], %[c10] \n"
" v_mul_f32 %[c11], %[scale_0], %[c11] \n"
" v_mul_f32 %[c12], %[scale_1], %[c12] \n"
" v_mul_f32 %[c13], %[scale_1], %[c13] \n"
" v_mul_f32 %[c14], %[scale_1], %[c14] \n"
" v_mul_f32 %[c15], %[scale_1], %[c15] \n" _UK_PK_CVT_(
"%[c0]", "%[c1]", "%[c0]") _UK_PK_CVT_("%[c2]", "%[c3]", "%[c1]")
_UK_PK_CVT_("%[c4]", "%[c5]", "%[c2]") _UK_PK_CVT_("%[c6]", "%[c7]", "%[c3]") _UK_PK_CVT_(
"%[c8]", "%[c9]", "%[c4]") _UK_PK_CVT_("%[c10]", "%[c11]", "%[c5]")
_UK_PK_CVT_("%[c12]", "%[c13]", "%[c6]") _UK_PK_CVT_(
"%[c14]",
"%[c15]",
"%[c7]") " ;------------------------------ \n"
" ds_write_b64 %[v_sfl_sst], [%[c0],%[c1]] offset:0 + %[shfl_base] "
" \n"
" ds_write_b64 %[v_sfl_sst], [%[c2],%[c3]] offset:4352 + %[shfl_base] "
" \n"
" ds_write_b64 %[v_sfl_sst], [%[c4],%[c5]] offset:2176 + %[shfl_base] "
" \n"
" ds_write_b64 %[v_sfl_sst], [%[c6],%[c7]] offset:6528 + %[shfl_base] "
" \n"
" s_waitcnt lgkmcnt(0) \n"
" s_barrier \n"
" ds_read_b32 %[c0], %[v_sfl_sld] offset:0 + %[shfl_base] "
" \n"
" ds_read_b32 %[c1], %[v_sfl_sld] offset:32 + %[shfl_base] "
" \n"
" ds_read_b32 %[c2], %[v_sfl_sld] offset:64 + %[shfl_base] "
" \n"
" ds_read_b32 %[c3], %[v_sfl_sld] offset:96 + %[shfl_base] "
" \n"
" ds_read_b32 %[c4], %[v_sfl_sld] offset:4352 + %[shfl_base] "
" \n"
" ds_read_b32 %[c5], %[v_sfl_sld] offset:4384 + %[shfl_base] "
" \n"
" ds_read_b32 %[c6], %[v_sfl_sld] offset:4416 + %[shfl_base] "
" \n"
" ds_read_b32 %[c7], %[v_sfl_sld] offset:4448 + %[shfl_base] "
" \n"
" s_waitcnt lgkmcnt(0) \n"
" s_mov_b64 exec, %[s_execflag_0] "
"\n" _UK_ATOMIC_ADD_ " %[v_os_o0], %[c0], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_1] "
"\n" _UK_ATOMIC_ADD_ " %[v_os_o1], %[c1], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_2] "
"\n" _UK_ATOMIC_ADD_ " %[v_os_o2], %[c2], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_3] "
"\n" _UK_ATOMIC_ADD_ " %[v_os_o3], %[c3], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_4] "
"\n" _UK_ATOMIC_ADD_ " %[v_os_o4], %[c4], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_5] "
"\n" _UK_ATOMIC_ADD_ " %[v_os_o5], %[c5], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_6] "
"\n" _UK_ATOMIC_ADD_ " %[v_os_o6], %[c6], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_7] "
"\n" _UK_ATOMIC_ADD_ " %[v_os_o7], %[c7], s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_sub_i32 %[s_loop_cnt], %[s_loop_cnt], 1 ; k-- \n"
" s_cmp_gt_i32 %[s_loop_cnt] 0 \n"
" s_cbranch_scc0 L_end%= \n"
" s_cmp_gt_i32 %[s_loop_cnt] 1 ; move b with cond \n"
" s_cselect_b32 s86, %[s_tile_os_b], 0 \n"
" s_add_u32 s12, s86, s12 \n"
" s_addc_u32 s13, 0, s13 \n"
" s_add_u32 s8, %[s_tile_os_o], s8 \n"
" s_addc_u32 s9, 0, s9 \n"
" s_waitcnt vmcnt(32) \n"
" s_barrier \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[128:129], v[128:129], 0 \n"
" buffer_load_dwordx4 acc[0:3], %[v_os_b0], s[12:15], 0 offen "
"\n" _UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[130:131], "
"v[130:131], [%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[132:133], v[132:133], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[134:135], v[134:135], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[136:137], v[136:137], "
"[%[c16],%[c17],%[c18],%[c19]] \n"
" buffer_load_dwordx4 acc[4:7], %[v_os_b0], s[12:15], 0 offen "
"offset:1024 \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[138:139], v[138:139], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[140:141], v[140:141], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[142:143], v[142:143], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[128:129], v[192:193], 0 \n"
" buffer_load_dwordx4 acc[8:11], %[v_os_b0], s[12:15], 0 offen "
"offset:2048 \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[130:131], v[194:195], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[132:133], v[196:197], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[134:135], v[198:199], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[136:137], v[200:201], "
"[%[c20],%[c21],%[c22],%[c23]] \n"
" buffer_load_dwordx4 acc[12:15], %[v_os_b0], s[12:15], 0 offen "
"offset:3072 \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[138:139], v[202:203], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[140:141], v[204:205], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[142:143], v[206:207], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[144:145], v[128:129], 0 \n"
" buffer_load_dwordx4 acc[16:19], %[v_os_b1], s[12:15], 0 offen "
"\n" _UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[146:147], "
"v[130:131], [%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[148:149], v[132:133], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[150:151], v[134:135], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[152:153], v[136:137], "
"[%[c24],%[c25],%[c26],%[c27]] \n"
" buffer_load_dwordx4 acc[20:23], %[v_os_b1], s[12:15], 0 offen "
"offset:1024 \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[154:155], v[138:139], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[156:157], v[140:141], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[158:159], v[142:143], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[144:145], v[192:193], 0 \n"
" buffer_load_dwordx4 acc[24:27], %[v_os_b1], s[12:15], 0 offen "
"offset:2048 \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[146:147], v[194:195], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[148:149], v[196:197], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[150:151], v[198:199], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[152:153], v[200:201], "
"[%[c28],%[c29],%[c30],%[c31]] \n"
" buffer_load_dwordx4 acc[28:31], %[v_os_b1], s[12:15], 0 offen "
"offset:3072 \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[154:155], v[202:203], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[156:157], v[204:205], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[158:159], v[206:207], "
"[%[c28],%[c29],%[c30],%[c31]] \n"
" s_waitcnt vmcnt(32) \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[160:161], v[144:145], "
"[%[c16],%[c17],%[c18],%[c19]] \n"
" buffer_load_dwordx4 acc[32:35], %[v_os_b2], s[12:15], 0 offen "
"\n" _UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[162:163], "
"v[146:147], [%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[164:165], v[148:149], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[166:167], v[150:151], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[168:169], v[152:153], "
"[%[c16],%[c17],%[c18],%[c19]] \n"
" buffer_load_dwordx4 acc[36:39], %[v_os_b2], s[12:15], 0 offen "
"offset:1024 \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[170:171], v[154:155], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[172:173], v[156:157], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[174:175], v[158:159], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[160:161], v[208:209], "
"[%[c20],%[c21],%[c22],%[c23]] \n"
" buffer_load_dwordx4 acc[40:43], %[v_os_b2], s[12:15], 0 offen "
"offset:2048 \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[162:163], v[210:211], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[164:165], v[212:213], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[166:167], v[214:215], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[168:169], v[216:217], "
"[%[c20],%[c21],%[c22],%[c23]] \n"
" buffer_load_dwordx4 acc[44:47], %[v_os_b2], s[12:15], 0 offen "
"offset:3072 \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[170:171], v[218:219], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[172:173], v[220:221], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[174:175], v[222:223], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[176:177], v[144:145], "
"[%[c24],%[c25],%[c26],%[c27]] \n"
" buffer_load_dwordx4 acc[48:51], %[v_os_b3], s[12:15], 0 offen "
"\n" _UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[178:179], "
"v[146:147], [%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[180:181], v[148:149], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[182:183], v[150:151], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[184:185], v[152:153], "
"[%[c24],%[c25],%[c26],%[c27]] \n"
" buffer_load_dwordx4 acc[52:55], %[v_os_b3], s[12:15], 0 offen "
"offset:1024 \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[186:187], v[154:155], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[188:189], v[156:157], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[190:191], v[158:159], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[176:177], v[208:209], "
"[%[c28],%[c29],%[c30],%[c31]] \n"
" buffer_load_dwordx4 acc[56:59], %[v_os_b3], s[12:15], 0 offen "
"offset:2048 \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[178:179], v[210:211], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[180:181], v[212:213], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[182:183], v[214:215], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[184:185], v[216:217], "
"[%[c28],%[c29],%[c30],%[c31]] \n"
" buffer_load_dwordx4 acc[60:63], %[v_os_b3], s[12:15], 0 offen "
"offset:3072 \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[186:187], v[218:219], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[188:189], v[220:221], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[190:191], v[222:223], "
"[%[c28],%[c29],%[c30],%[c31]] \n"
" s_waitcnt vmcnt(32) \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[192:193], v[160:161], "
"[%[c16],%[c17],%[c18],%[c19]] \n"
" buffer_load_dwordx4 acc[64:67], %[v_os_b4], s[12:15], 0 offen "
"\n" _UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[194:195], "
"v[162:163], [%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[196:197], v[164:165], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[198:199], v[166:167], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[200:201], v[168:169], "
"[%[c16],%[c17],%[c18],%[c19]] \n"
" buffer_load_dwordx4 acc[68:71], %[v_os_b4], s[12:15], 0 offen "
"offset:1024 \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[202:203], v[170:171], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[204:205], v[172:173], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[206:207], v[174:175], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[192:193], v[224:225], "
"[%[c20],%[c21],%[c22],%[c23]] \n"
" buffer_load_dwordx4 acc[72:75], %[v_os_b4], s[12:15], 0 offen "
"offset:2048 \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[194:195], v[226:227], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[196:197], v[228:229], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[198:199], v[230:231], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[200:201], v[232:233], "
"[%[c20],%[c21],%[c22],%[c23]] \n"
" buffer_load_dwordx4 acc[76:79], %[v_os_b4], s[12:15], 0 offen "
"offset:3072 \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[202:203], v[234:235], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[204:205], v[236:237], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[206:207], v[238:239], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[208:209], v[160:161], "
"[%[c24],%[c25],%[c26],%[c27]] \n"
" buffer_load_dwordx4 acc[80:83], %[v_os_b5], s[12:15], 0 offen "
"\n" _UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[210:211], "
"v[162:163], [%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[212:213], v[164:165], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[214:215], v[166:167], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[216:217], v[168:169], "
"[%[c24],%[c25],%[c26],%[c27]] \n"
" buffer_load_dwordx4 acc[84:87], %[v_os_b5], s[12:15], 0 offen "
"offset:1024 \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[218:219], v[170:171], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[220:221], v[172:173], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[222:223], v[174:175], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[208:209], v[224:225], "
"[%[c28],%[c29],%[c30],%[c31]] \n"
" buffer_load_dwordx4 acc[88:91], %[v_os_b5], s[12:15], 0 offen "
"offset:2048 \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[210:211], v[226:227], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[212:213], v[228:229], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[214:215], v[230:231], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[216:217], v[232:233], "
"[%[c28],%[c29],%[c30],%[c31]] \n"
" buffer_load_dwordx4 acc[92:95], %[v_os_b5], s[12:15], 0 offen "
"offset:3072 \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[218:219], v[234:235], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[220:221], v[236:237], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[222:223], v[238:239], "
"[%[c28],%[c29],%[c30],%[c31]] \n"
" s_waitcnt vmcnt(32) \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[224:225], v[176:177], "
"[%[c16],%[c17],%[c18],%[c19]] \n"
" buffer_load_dwordx4 acc[96:99], %[v_os_b6], s[12:15], 0 offen "
"\n" _UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[226:227], "
"v[178:179], [%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[228:229], v[180:181], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[230:231], v[182:183], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[232:233], v[184:185], "
"[%[c16],%[c17],%[c18],%[c19]] \n"
" buffer_load_dwordx4 acc[100:103], %[v_os_b6], s[12:15], 0 offen "
"offset:1024 \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[234:235], v[186:187], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[236:237], v[188:189], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c16],%[c17],%[c18],%[c19]], acc[238:239], v[190:191], "
"[%[c16],%[c17],%[c18],%[c19]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[224:225], v[240:241], "
"[%[c20],%[c21],%[c22],%[c23]] \n"
" buffer_load_dwordx4 acc[104:107], %[v_os_b6], s[12:15], 0 offen "
"offset:2048 \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[226:227], v[242:243], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[228:229], v[244:245], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[230:231], v[246:247], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[232:233], v[248:249], "
"[%[c20],%[c21],%[c22],%[c23]] \n"
" buffer_load_dwordx4 acc[108:111], %[v_os_b6], s[12:15], 0 offen "
"offset:3072 \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[234:235], v[250:251], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[236:237], v[252:253], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c20],%[c21],%[c22],%[c23]], acc[238:239], v[254:255], "
"[%[c20],%[c21],%[c22],%[c23]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[240:241], v[176:177], "
"[%[c24],%[c25],%[c26],%[c27]] \n"
" buffer_load_dwordx4 acc[112:115], %[v_os_b7], s[12:15], 0 offen "
"\n" _UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[242:243], "
"v[178:179], [%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[244:245], v[180:181], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[246:247], v[182:183], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[248:249], v[184:185], "
"[%[c24],%[c25],%[c26],%[c27]] \n"
" buffer_load_dwordx4 acc[116:119], %[v_os_b7], s[12:15], 0 offen "
"offset:1024 \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[250:251], v[186:187], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[252:253], v[188:189], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c24],%[c25],%[c26],%[c27]], acc[254:255], v[190:191], "
"[%[c24],%[c25],%[c26],%[c27]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[240:241], v[240:241], "
"[%[c28],%[c29],%[c30],%[c31]] \n"
" buffer_load_dwordx4 acc[120:123], %[v_os_b7], s[12:15], 0 offen "
"offset:2048 \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[242:243], v[242:243], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[244:245], v[244:245], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[246:247], v[246:247], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[248:249], v[248:249], "
"[%[c28],%[c29],%[c30],%[c31]] \n"
" buffer_load_dwordx4 acc[124:127], %[v_os_b7], s[12:15], 0 offen "
"offset:3072 \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[250:251], v[250:251], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[252:253], v[252:253], "
"[%[c28],%[c29],%[c30],%[c31]] \n" _UK_MFMA_
" [%[c28],%[c29],%[c30],%[c31]], acc[254:255], v[254:255], "
"[%[c28],%[c29],%[c30],%[c31]]\n"
" v_mul_f32 %[c16], %[scale_0], %[c16] \n"
" v_mul_f32 %[c17], %[scale_0], %[c17] \n"
" v_mul_f32 %[c18], %[scale_0], %[c18] \n"
" v_mul_f32 %[c19], %[scale_0], %[c19] \n"
" v_mul_f32 %[c20], %[scale_1], %[c20] \n"
" v_mul_f32 %[c21], %[scale_1], %[c21] \n"
" v_mul_f32 %[c22], %[scale_1], %[c22] \n"
" v_mul_f32 %[c23], %[scale_1], %[c23] \n"
" v_mul_f32 %[c24], %[scale_0], %[c24] \n"
" v_mul_f32 %[c25], %[scale_0], %[c25] \n"
" v_mul_f32 %[c26], %[scale_0], %[c26] \n"
" v_mul_f32 %[c27], %[scale_0], %[c27] \n"
" v_mul_f32 %[c28], %[scale_1], %[c28] \n"
" v_mul_f32 %[c29], %[scale_1], %[c29] \n"
" v_mul_f32 %[c30], %[scale_1], %[c30] \n"
" v_mul_f32 %[c31], %[scale_1], %[c31] \n"
_UK_PK_CVT_("%[c16]", "%[c17]", "%[c16]")
_UK_PK_CVT_("%[c18]", "%[c19]", "%[c17]")
_UK_PK_CVT_("%[c20]", "%[c21]", "%[c18]")
_UK_PK_CVT_("%[c22]", "%[c23]", "%[c19]")
_UK_PK_CVT_("%[c24]", "%[c25]", "%[c20]")
_UK_PK_CVT_("%[c26]", "%[c27]", "%[c21]")
_UK_PK_CVT_("%[c28]", "%[c29]", "%[c22]")
_UK_PK_CVT_("%[c30]", "%[c31]", "%[c23]")
_UK_PK_CVT_("%[c16]", "%[c17]", "%[c16]") _UK_PK_CVT_("%[c18]", "%[c19]", "%[c17]") _UK_PK_CVT_(
"%[c20]", "%[c21]", "%[c18]") _UK_PK_CVT_("%[c22]", "%[c23]", "%[c19]")
_UK_PK_CVT_("%[c24]", "%[c25]", "%[c20]") _UK_PK_CVT_(
"%[c26]", "%[c27]", "%[c21]") _UK_PK_CVT_("%[c28]",
"%[c29]",
"%[c22]") _UK_PK_CVT_("%[c30]",
"%[c31]",
"%[c23]")
" ;------------------------------ \n"
" ds_write_b64 %[v_sfl_sst], [%[c16],%[c17]] offset:0 + %[shfl_base] \n"
" ds_write_b64 %[v_sfl_sst], [%[c18],%[c19]] offset:4352 + %[shfl_base] \n"
" ds_write_b64 %[v_sfl_sst], [%[c20],%[c21]] offset:2176 + %[shfl_base] \n"
" ds_write_b64 %[v_sfl_sst], [%[c22],%[c23]] offset:6528 + %[shfl_base] \n"
" s_waitcnt lgkmcnt(0) \n"
" s_barrier \n"
" ds_read_b32 %[c16], %[v_sfl_sld] offset:0 + %[shfl_base] \n"
" ds_read_b32 %[c17], %[v_sfl_sld] offset:32 + %[shfl_base] \n"
" ds_read_b32 %[c18], %[v_sfl_sld] offset:64 + %[shfl_base] \n"
" ds_read_b32 %[c19], %[v_sfl_sld] offset:96 + %[shfl_base] \n"
" ds_read_b32 %[c20], %[v_sfl_sld] offset:4352 + %[shfl_base] \n"
" ds_read_b32 %[c21], %[v_sfl_sld] offset:4384 + %[shfl_base] \n"
" ds_read_b32 %[c22], %[v_sfl_sld] offset:4416 + %[shfl_base] \n"
" ds_read_b32 %[c23], %[v_sfl_sld] offset:4448 + %[shfl_base] \n"
" s_waitcnt lgkmcnt(0) \n"
" s_mov_b64 exec, %[s_execflag_0] \n"
_UK_ATOMIC_ADD_ " %[v_os_o0], %[c16], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_1] \n"
_UK_ATOMIC_ADD_ " %[v_os_o1], %[c17], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_2] \n"
_UK_ATOMIC_ADD_ " %[v_os_o2], %[c18], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_3] \n"
_UK_ATOMIC_ADD_ " %[v_os_o3], %[c19], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_4] \n"
_UK_ATOMIC_ADD_ " %[v_os_o4], %[c20], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_5] \n"
_UK_ATOMIC_ADD_ " %[v_os_o5], %[c21], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_6] \n"
_UK_ATOMIC_ADD_ " %[v_os_o6], %[c22], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_7] \n"
_UK_ATOMIC_ADD_ " %[v_os_o7], %[c23], s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_sub_i32 %[s_loop_cnt], %[s_loop_cnt], 1 ; k-- \n"
" s_cmp_gt_i32 %[s_loop_cnt] 0 \n"
" s_cbranch_scc0 L_end%= \n"
" s_cmp_gt_i32 %[s_loop_cnt] 1 ; move b with cond \n"
" s_cselect_b32 s86, %[s_tile_os_b], 0 \n"
" s_add_u32 s12, s86, s12 \n"
" s_addc_u32 s13, 0, s13 \n"
" s_add_u32 s8, %[s_tile_os_o], s8 \n"
" s_addc_u32 s9, 0, s9 \n"
" s_branch L_start%= \n"
"L_end%=: \n"
" ;------------------------------ \n"
" ds_write_b64 %[v_sfl_sst], [%[c16],%[c17]] offset:0 + %[shfl_base] \n"
" ds_write_b64 %[v_sfl_sst], [%[c18],%[c19]] offset:4352 + %[shfl_base] \n"
" ds_write_b64 %[v_sfl_sst], [%[c20],%[c21]] offset:2176 + %[shfl_base] \n"
" ds_write_b64 %[v_sfl_sst], [%[c22],%[c23]] offset:6528 + %[shfl_base] \n"
" s_waitcnt lgkmcnt(0) \n"
" s_barrier \n"
" ds_read_b32 %[c16], %[v_sfl_sld] offset:0 + %[shfl_base] \n"
" ds_read_b32 %[c17], %[v_sfl_sld] offset:32 + %[shfl_base] \n"
" ds_read_b32 %[c18], %[v_sfl_sld] offset:64 + %[shfl_base] \n"
" ds_read_b32 %[c19], %[v_sfl_sld] offset:96 + %[shfl_base] \n"
" ds_read_b32 %[c20], %[v_sfl_sld] offset:4352 + %[shfl_base] \n"
" ds_read_b32 %[c21], %[v_sfl_sld] offset:4384 + %[shfl_base] \n"
" ds_read_b32 %[c22], %[v_sfl_sld] offset:4416 + %[shfl_base] \n"
" ds_read_b32 %[c23], %[v_sfl_sld] offset:4448 + %[shfl_base] \n"
" s_waitcnt lgkmcnt(0) \n"
" s_mov_b64 exec, %[s_execflag_0] \n" _UK_ATOMIC_ADD_
" %[v_os_o0], %[c16], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_1] \n" _UK_ATOMIC_ADD_
" %[v_os_o1], %[c17], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_2] \n" _UK_ATOMIC_ADD_
" %[v_os_o2], %[c18], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_3] \n" _UK_ATOMIC_ADD_
" %[v_os_o3], %[c19], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_4] \n" _UK_ATOMIC_ADD_
" %[v_os_o4], %[c20], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_5] \n" _UK_ATOMIC_ADD_
" %[v_os_o5], %[c21], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_6] \n" _UK_ATOMIC_ADD_
" %[v_os_o6], %[c22], s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_7] \n" _UK_ATOMIC_ADD_
" %[v_os_o7], %[c23], s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_sub_i32 %[s_loop_cnt], %[s_loop_cnt], 1 ; k-- \n"
" s_cmp_gt_i32 %[s_loop_cnt] 0 \n"
" s_cbranch_scc0 L_end%= \n"
" s_cmp_gt_i32 %[s_loop_cnt] 1 ; move b with cond \n"
" s_cselect_b32 s86, %[s_tile_os_b], 0 \n"
" s_add_u32 s12, s86, s12 \n"
" s_addc_u32 s13, 0, s13 \n"
" s_add_u32 s8, %[s_tile_os_o], s8 \n"
" s_addc_u32 s9, 0, s9 \n"
" s_branch L_start%= \n"
"L_end%=: \n"
#undef _UK_MFMA_
#undef _UK_PK_CVT_
......
include/ck_tile/ops/flatmm/block/uk/flatmm_sn_uk_gfx9_32x128x512_1x4x1_16x16x16_itl.inc
View file @ 2a768425
......@@ -3,706 +3,767 @@
#endif
#if CK_TILE_FLATMM_UK_MFMA == CK_TILE_FLATMM_UK_MFMA_BF16
# define _UK_MFMA_ "v_mfma_f32_16x16x16_bf16"
#define _UK_MFMA_ "v_mfma_f32_16x16x16_bf16"
# define _UK_PK_CVT_(x0_, x1_, y_) \
" v_cmp_u_f32 s[36:37], " x0_ ", " x0_ " \n" \
" v_add3_u32 v50, " x0_ ", %[v_nan_lo], 1 \n" \
" v_cndmask_b32 v54, v50, %[v_nan_hi], s[36:37] \n" \
" v_cmp_u_f32 s[36:37], " x1_ ", " x1_ " \n" \
" v_add3_u32 v50, " x1_ ", %[v_nan_lo], 1 \n" \
" v_cndmask_b32 v55, v50, %[v_nan_hi], s[36:37] \n" \
" v_perm_b32 " y_ ", v55, v54, s52 \n"
#define _UK_PK_CVT_(x0_, x1_, y_) \
" v_cmp_u_f32 s[36:37], " x0_ ", " x0_ " \n" \
" v_add3_u32 v50, " x0_ ", %[v_nan_lo], 1 \n" \
" v_cndmask_b32 v54, v50, %[v_nan_hi], s[36:37] \n" \
" v_cmp_u_f32 s[36:37], " x1_ ", " x1_ " \n" \
" v_add3_u32 v50, " x1_ ", %[v_nan_lo], 1 \n" \
" v_cndmask_b32 v55, v50, %[v_nan_hi], s[36:37] \n" \
" v_perm_b32 " y_ ", v55, v54, s52 \n"
# define _UK_ATOMIC_ADD_ "global_atomic_pk_add_bf16"
#define _UK_ATOMIC_ADD_ "global_atomic_pk_add_bf16"
#elif CK_TILE_FLATMM_UK_MFMA == CK_TILE_FLATMM_UK_MFMA_FP16
#define _UK_MFMA_ "v_mfma_f32_16x16x16_f16"
# define _UK_PK_CVT_(x0_, x1_, y_) \
" v_cvt_f16_f32 v54, " x0_ " \n" \
" v_cvt_f16_f32 v55, " x1_ " \n" \
" v_pack_b32_f16 " y_ ", v54, v55 \n"
#define _UK_PK_CVT_(x0_, x1_, y_) \
" v_cvt_f16_f32 v54, " x0_ " \n" \
" v_cvt_f16_f32 v55, " x1_ " \n" \
" v_pack_b32_f16 " y_ ", v54, v55 \n"
# define _UK_ATOMIC_ADD_ "global_atomic_pk_add_f16"
#define _UK_ATOMIC_ADD_ "global_atomic_pk_add_f16"
#endif
";-------------------------------------------------------------\n"
" s_mov_b32 s52, 0x07060302 ; v_perm\n"
" s_mov_b64 s[38:39], exec ; save current exec\n"
" s_mov_b32 s8, %[s_res_o0] \n"
" s_mov_b32 s9, %[s_res_o1] \n"
" s_mov_b32 s12, %[s_res_b0] \n"
" s_mov_b32 s13, %[s_res_b1] \n"
" s_mov_b32 s14, %[s_res_b2] \n"
" s_mov_b32 s15, %[s_res_b3] \n"
" s_mov_b32 s59, 0 \n"
" ds_read_b64 v[128:129], %[v_sld_y_os] offset:0 + %[sld_a_base] \n"
" ds_read_b64 v[130:131], %[v_sld_y_os] offset:128 + %[sld_a_base] \n"
" ds_read_b64 v[132:133], %[v_sld_y_os] offset:1024 + %[sld_a_base] \n"
" ds_read_b64 v[134:135], %[v_sld_y_os] offset:1152 + %[sld_a_base] \n"
" ds_read_b64 v[136:137], %[v_sld_y_os] offset:2048 + %[sld_a_base] \n"
" ds_read_b64 v[138:139], %[v_sld_y_os] offset:2176 + %[sld_a_base] \n"
" ds_read_b64 v[140:141], %[v_sld_y_os] offset:3072 + %[sld_a_base] \n"
" ds_read_b64 v[142:143], %[v_sld_y_os] offset:3200 + %[sld_a_base] \n"
" ds_read_b64 v[144:145], %[v_sld_y_os] offset:4096 + %[sld_a_base] \n"
" ds_read_b64 v[146:147], %[v_sld_y_os] offset:4224 + %[sld_a_base] \n"
" ds_read_b64 v[148:149], %[v_sld_y_os] offset:5120 + %[sld_a_base] \n"
" ds_read_b64 v[150:151], %[v_sld_y_os] offset:5248 + %[sld_a_base] \n"
" ds_read_b64 v[152:153], %[v_sld_y_os] offset:6144 + %[sld_a_base] \n"
" ds_read_b64 v[154:155], %[v_sld_y_os] offset:6272 + %[sld_a_base] \n"
" ds_read_b64 v[156:157], %[v_sld_y_os] offset:7168 + %[sld_a_base] \n"
" ds_read_b64 v[158:159], %[v_sld_y_os] offset:7296 + %[sld_a_base] \n"
" ds_read_b64 v[160:161], %[v_sld_y_os] offset:8192 + %[sld_a_base] \n"
" ds_read_b64 v[162:163], %[v_sld_y_os] offset:8320 + %[sld_a_base] \n"
" ds_read_b64 v[164:165], %[v_sld_y_os] offset:9216 + %[sld_a_base] \n"
" ds_read_b64 v[166:167], %[v_sld_y_os] offset:9344 + %[sld_a_base] \n"
" ds_read_b64 v[168:169], %[v_sld_y_os] offset:10240 + %[sld_a_base] \n"
" ds_read_b64 v[170:171], %[v_sld_y_os] offset:10368 + %[sld_a_base] \n"
" ds_read_b64 v[172:173], %[v_sld_y_os] offset:11264 + %[sld_a_base] \n"
" ds_read_b64 v[174:175], %[v_sld_y_os] offset:11392 + %[sld_a_base] \n"
" ds_read_b64 v[176:177], %[v_sld_y_os] offset:12288 + %[sld_a_base] \n"
" ds_read_b64 v[178:179], %[v_sld_y_os] offset:12416 + %[sld_a_base] \n"
" ds_read_b64 v[180:181], %[v_sld_y_os] offset:13312 + %[sld_a_base] \n"
" ds_read_b64 v[182:183], %[v_sld_y_os] offset:13440 + %[sld_a_base] \n"
" ds_read_b64 v[184:185], %[v_sld_y_os] offset:14336 + %[sld_a_base] \n"
" ds_read_b64 v[186:187], %[v_sld_y_os] offset:14464 + %[sld_a_base] \n"
" ds_read_b64 v[188:189], %[v_sld_y_os] offset:15360 + %[sld_a_base] \n"
" ds_read_b64 v[190:191], %[v_sld_y_os] offset:15488 + %[sld_a_base] \n"
" ds_read_b64 v[192:193], %[v_sld_y_os] offset:16384 + %[sld_a_base] \n"
" ds_read_b64 v[194:195], %[v_sld_y_os] offset:16512 + %[sld_a_base] \n"
" ds_read_b64 v[196:197], %[v_sld_y_os] offset:17408 + %[sld_a_base] \n"
" ds_read_b64 v[198:199], %[v_sld_y_os] offset:17536 + %[sld_a_base] \n"
" ds_read_b64 v[200:201], %[v_sld_y_os] offset:18432 + %[sld_a_base] \n"
" ds_read_b64 v[202:203], %[v_sld_y_os] offset:18560 + %[sld_a_base] \n"
" ds_read_b64 v[204:205], %[v_sld_y_os] offset:19456 + %[sld_a_base] \n"
" ds_read_b64 v[206:207], %[v_sld_y_os] offset:19584 + %[sld_a_base] \n"
" ds_read_b64 v[208:209], %[v_sld_y_os] offset:20480 + %[sld_a_base] \n"
" ds_read_b64 v[210:211], %[v_sld_y_os] offset:20608 + %[sld_a_base] \n"
" ds_read_b64 v[212:213], %[v_sld_y_os] offset:21504 + %[sld_a_base] \n"
" ds_read_b64 v[214:215], %[v_sld_y_os] offset:21632 + %[sld_a_base] \n"
" ds_read_b64 v[216:217], %[v_sld_y_os] offset:22528 + %[sld_a_base] \n"
" ds_read_b64 v[218:219], %[v_sld_y_os] offset:22656 + %[sld_a_base] \n"
" ds_read_b64 v[220:221], %[v_sld_y_os] offset:23552 + %[sld_a_base] \n"
" ds_read_b64 v[222:223], %[v_sld_y_os] offset:23680 + %[sld_a_base] \n"
" ds_read_b64 v[224:225], %[v_sld_y_os] offset:24576 + %[sld_a_base] \n"
" ds_read_b64 v[226:227], %[v_sld_y_os] offset:24704 + %[sld_a_base] \n"
" ds_read_b64 v[228:229], %[v_sld_y_os] offset:25600 + %[sld_a_base] \n"
" ds_read_b64 v[230:231], %[v_sld_y_os] offset:25728 + %[sld_a_base] \n"
" ds_read_b64 v[232:233], %[v_sld_y_os] offset:26624 + %[sld_a_base] \n"
" ds_read_b64 v[234:235], %[v_sld_y_os] offset:26752 + %[sld_a_base] \n"
" ds_read_b64 v[236:237], %[v_sld_y_os] offset:27648 + %[sld_a_base] \n"
" ds_read_b64 v[238:239], %[v_sld_y_os] offset:27776 + %[sld_a_base] \n"
" ds_read_b64 v[240:241], %[v_sld_y_os] offset:28672 + %[sld_a_base] \n"
" ds_read_b64 v[242:243], %[v_sld_y_os] offset:28800 + %[sld_a_base] \n"
" ds_read_b64 v[244:245], %[v_sld_y_os] offset:29696 + %[sld_a_base] \n"
" ds_read_b64 v[246:247], %[v_sld_y_os] offset:29824 + %[sld_a_base] \n"
" ds_read_b64 v[248:249], %[v_sld_y_os] offset:30720 + %[sld_a_base] \n"
" ds_read_b64 v[250:251], %[v_sld_y_os] offset:30848 + %[sld_a_base] \n"
" ds_read_b64 v[252:253], %[v_sld_y_os] offset:31744 + %[sld_a_base] \n"
" ds_read_b64 v[254:255], %[v_sld_y_os] offset:31872 + %[sld_a_base] \n"
" s_waitcnt 0 \n"
" buffer_load_dwordx4 acc[0:3], %[v_os_b0], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[4:7], %[v_os_b0], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[8:11], %[v_os_b0], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[12:15], %[v_os_b0], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[16:19], %[v_os_b1], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[20:23], %[v_os_b1], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[24:27], %[v_os_b1], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[28:31], %[v_os_b1], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[32:35], %[v_os_b2], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[36:39], %[v_os_b2], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[40:43], %[v_os_b2], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[44:47], %[v_os_b2], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[48:51], %[v_os_b3], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[52:55], %[v_os_b3], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[56:59], %[v_os_b3], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[60:63], %[v_os_b3], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[64:67], %[v_os_b4], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[68:71], %[v_os_b4], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[72:75], %[v_os_b4], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[76:79], %[v_os_b4], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[80:83], %[v_os_b5], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[84:87], %[v_os_b5], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[88:91], %[v_os_b5], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[92:95], %[v_os_b5], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[96:99], %[v_os_b6], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[100:103], %[v_os_b6], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[104:107], %[v_os_b6], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[108:111], %[v_os_b6], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[112:115], %[v_os_b7], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[116:119], %[v_os_b7], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[120:123], %[v_os_b7], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[124:127], %[v_os_b7], s[12:15], 0 offen offset:3072 \n"
" s_add_u32 s12, %[s_tile_os_b], s12 \n"
" s_addc_u32 s13, 0, s13 \n"
" v_mov_b32 v64, 0 \n"
" v_mov_b32 v80, 0 \n"
" v_mov_b32 v65, 0 \n"
" v_mov_b32 v81, 0 \n"
" v_mov_b32 v66, 0 \n"
" v_mov_b32 v82, 0 \n"
" v_mov_b32 v67, 0 \n"
" v_mov_b32 v83, 0 \n"
" v_mov_b32 v68, 0 \n"
" v_mov_b32 v84, 0 \n"
" v_mov_b32 v69, 0 \n"
" v_mov_b32 v85, 0 \n"
" v_mov_b32 v70, 0 \n"
" v_mov_b32 v86, 0 \n"
" v_mov_b32 v71, 0 \n"
" v_mov_b32 v87, 0 \n"
" ds_write_b64 %[v_sfl_sst], [%[c0],%[c1]] offset:16640 \n"
" ds_write_b64 %[v_sfl_sst], [%[c2],%[c3]] offset:20992 \n"
" ds_write_b64 %[v_sfl_sst], [%[c4],%[c5]] offset:18816 \n"
" ds_write_b64 %[v_sfl_sst], [%[c6],%[c7]] offset:23168 \n"
" s_mov_b32 s80, 0 \n"
" s_waitcnt vmcnt(24) \n"
"label_0AA6: \n"
" s_waitcnt vmcnt(30) & lgkmcnt(0) \n"
" s_barrier \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[0:1], v[128:129], 0 \n"
" ds_read_b32 v10, %[v_sfl_sld] offset:16640 \n"
" ds_read_b32 v11, %[v_sfl_sld] offset:16672 \n"
" ds_write_b64 %[v_sfl_sst], [%[c16],%[c17]] offset:25344 \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[2:3], v[130:131], v[64:67] \n"
" buffer_load_dwordx4 acc[128:131], %[v_os_b0], s[12:15], 0 offen \n"
" ds_write_b64 %[v_sfl_sst], [%[c18],%[c19]] offset:29696 \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[4:5], v[132:133], v[64:67] \n"
" ds_read_b32 v12, %[v_sfl_sld] offset:16704 \n"
" ds_read_b32 v13, %[v_sfl_sld] offset:16736 \n"
" ds_write_b64 %[v_sfl_sst], [%[c20],%[c21]] offset:27520 \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[6:7], v[134:135], v[64:67] \n"
" ds_write_b64 %[v_sfl_sst], [%[c22],%[c23]] offset:31872 \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[8:9], v[136:137], v[64:67] \n"
" ds_read_b32 v14, %[v_sfl_sld] offset:20992 \n"
" ds_read_b32 v15, %[v_sfl_sld] offset:21024 \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[10:11], v[138:139], v[64:67] \n"
" buffer_load_dwordx4 acc[132:135], %[v_os_b0], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[12:13], v[140:141], v[64:67] \n"
" ds_read_b32 v16, %[v_sfl_sld] offset:21056 \n"
" ds_read_b32 v17, %[v_sfl_sld] offset:21088 \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[14:15], v[142:143], v[64:67] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[0:1], v[192:193], 0 \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[2:3], v[194:195], v[68:71] \n"
" buffer_load_dwordx4 acc[136:139], %[v_os_b0], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[4:5], v[196:197], v[68:71] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[6:7], v[198:199], v[68:71] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[8:9], v[200:201], v[68:71] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[10:11], v[202:203], v[68:71] \n"
" buffer_load_dwordx4 acc[140:143], %[v_os_b0], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[12:13], v[204:205], v[68:71] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[14:15], v[206:207], v[68:71] \n"
" s_waitcnt lgkmcnt(0) \n"
" s_mov_b64 exec, %[s_execflag_0] \n"
_UK_ATOMIC_ADD_ " %[v_os_o0], v10, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[16:17], v[128:129], 0 \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[18:19], v[130:131], v[72:75] \n"
" buffer_load_dwordx4 acc[144:147], %[v_os_b1], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[20:21], v[132:133], v[72:75] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[22:23], v[134:135], v[72:75] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[24:25], v[136:137], v[72:75] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[26:27], v[138:139], v[72:75] \n"
" buffer_load_dwordx4 acc[148:151], %[v_os_b1], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[28:29], v[140:141], v[72:75] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[30:31], v[142:143], v[72:75] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[16:17], v[192:193], 0 \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[18:19], v[194:195], v[76:79] \n"
" buffer_load_dwordx4 acc[152:155], %[v_os_b1], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[20:21], v[196:197], v[76:79] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[22:23], v[198:199], v[76:79] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[24:25], v[200:201], v[76:79] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[26:27], v[202:203], v[76:79] \n"
" buffer_load_dwordx4 acc[156:159], %[v_os_b1], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[28:29], v[204:205], v[76:79] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[30:31], v[206:207], v[76:79] \n"
" s_mov_b64 exec, %[s_execflag_1] \n"
_UK_ATOMIC_ADD_ " %[v_os_o1], v11, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_waitcnt vmcnt(30) \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[32:33], v[144:145], v[64:67] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[34:35], v[146:147], v[64:67] \n"
" buffer_load_dwordx4 acc[160:163], %[v_os_b2], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[36:37], v[148:149], v[64:67] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[38:39], v[150:151], v[64:67] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[40:41], v[152:153], v[64:67] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[42:43], v[154:155], v[64:67] \n"
" buffer_load_dwordx4 acc[164:167], %[v_os_b2], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[44:45], v[156:157], v[64:67] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[46:47], v[158:159], v[64:67] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[32:33], v[208:209], v[68:71] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[34:35], v[210:211], v[68:71] \n"
" buffer_load_dwordx4 acc[168:171], %[v_os_b2], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[36:37], v[212:213], v[68:71] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[38:39], v[214:215], v[68:71] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[40:41], v[216:217], v[68:71] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[42:43], v[218:219], v[68:71] \n"
" buffer_load_dwordx4 acc[172:175], %[v_os_b2], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[44:45], v[220:221], v[68:71] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[46:47], v[222:223], v[68:71] \n"
" s_mov_b64 exec, %[s_execflag_2] \n"
_UK_ATOMIC_ADD_ " %[v_os_o2], v12, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[48:49], v[144:145], v[72:75] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[50:51], v[146:147], v[72:75] \n"
" buffer_load_dwordx4 acc[176:179], %[v_os_b3], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[52:53], v[148:149], v[72:75] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[54:55], v[150:151], v[72:75] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[56:57], v[152:153], v[72:75] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[58:59], v[154:155], v[72:75] \n"
" buffer_load_dwordx4 acc[180:183], %[v_os_b3], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[60:61], v[156:157], v[72:75] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[62:63], v[158:159], v[72:75] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[48:49], v[208:209], v[76:79] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[50:51], v[210:211], v[76:79] \n"
" buffer_load_dwordx4 acc[184:187], %[v_os_b3], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[52:53], v[212:213], v[76:79] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[54:55], v[214:215], v[76:79] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[56:57], v[216:217], v[76:79] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[58:59], v[218:219], v[76:79] \n"
" buffer_load_dwordx4 acc[188:191], %[v_os_b3], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[60:61], v[220:221], v[76:79] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[62:63], v[222:223], v[76:79] \n"
" s_mov_b64 exec, %[s_execflag_3] \n"
_UK_ATOMIC_ADD_ " %[v_os_o3], v13, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_waitcnt vmcnt(30) \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[64:65], v[160:161], v[64:67] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[66:67], v[162:163], v[64:67] \n"
" buffer_load_dwordx4 acc[192:195], %[v_os_b4], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[68:69], v[164:165], v[64:67] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[70:71], v[166:167], v[64:67] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[72:73], v[168:169], v[64:67] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[74:75], v[170:171], v[64:67] \n"
" buffer_load_dwordx4 acc[196:199], %[v_os_b4], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[76:77], v[172:173], v[64:67] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[78:79], v[174:175], v[64:67] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[64:65], v[224:225], v[68:71] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[66:67], v[226:227], v[68:71] \n"
" buffer_load_dwordx4 acc[200:203], %[v_os_b4], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[68:69], v[228:229], v[68:71] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[70:71], v[230:231], v[68:71] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[72:73], v[232:233], v[68:71] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[74:75], v[234:235], v[68:71] \n"
" buffer_load_dwordx4 acc[204:207], %[v_os_b4], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[76:77], v[236:237], v[68:71] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[78:79], v[238:239], v[68:71] \n"
" s_mov_b64 exec, %[s_execflag_4] \n"
_UK_ATOMIC_ADD_ " %[v_os_o4], v14, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[80:81], v[160:161], v[72:75] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[82:83], v[162:163], v[72:75] \n"
" buffer_load_dwordx4 acc[208:211], %[v_os_b5], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[84:85], v[164:165], v[72:75] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[86:87], v[166:167], v[72:75] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[88:89], v[168:169], v[72:75] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[90:91], v[170:171], v[72:75] \n"
" buffer_load_dwordx4 acc[212:215], %[v_os_b5], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[92:93], v[172:173], v[72:75] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[94:95], v[174:175], v[72:75] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[80:81], v[224:225], v[76:79] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[82:83], v[226:227], v[76:79] \n"
" buffer_load_dwordx4 acc[216:219], %[v_os_b5], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[84:85], v[228:229], v[76:79] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[86:87], v[230:231], v[76:79] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[88:89], v[232:233], v[76:79] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[90:91], v[234:235], v[76:79] \n"
" buffer_load_dwordx4 acc[220:223], %[v_os_b5], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[92:93], v[236:237], v[76:79] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[94:95], v[238:239], v[76:79] \n"
" s_mov_b64 exec, %[s_execflag_5] \n"
_UK_ATOMIC_ADD_ " %[v_os_o5], v15, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_waitcnt vmcnt(30) \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[96:97], v[176:177], v[64:67] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[98:99], v[178:179], v[64:67] \n"
" buffer_load_dwordx4 acc[224:227], %[v_os_b6], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[100:101], v[180:181], v[64:67] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[102:103], v[182:183], v[64:67] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[104:105], v[184:185], v[64:67] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[106:107], v[186:187], v[64:67] \n"
" buffer_load_dwordx4 acc[228:231], %[v_os_b6], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[108:109], v[188:189], v[64:67] \n"
_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[110:111], v[190:191], v[64:67] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[96:97], v[240:241], v[68:71] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[98:99], v[242:243], v[68:71] \n"
" buffer_load_dwordx4 acc[232:235], %[v_os_b6], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[100:101], v[244:245], v[68:71] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[102:103], v[246:247], v[68:71] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[104:105], v[248:249], v[68:71] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[106:107], v[250:251], v[68:71] \n"
" buffer_load_dwordx4 acc[236:239], %[v_os_b6], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[108:109], v[252:253], v[68:71] \n"
_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[110:111], v[254:255], v[68:71] \n"
" s_mov_b64 exec, %[s_execflag_6] \n"
_UK_ATOMIC_ADD_ " %[v_os_o6], v16, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[112:113], v[176:177], v[72:75] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[114:115], v[178:179], v[72:75] \n"
" buffer_load_dwordx4 acc[240:243], %[v_os_b7], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[116:117], v[180:181], v[72:75] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[118:119], v[182:183], v[72:75] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[120:121], v[184:185], v[72:75] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[122:123], v[186:187], v[72:75] \n"
" buffer_load_dwordx4 acc[244:247], %[v_os_b7], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[124:125], v[188:189], v[72:75] \n"
_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[126:127], v[190:191], v[72:75] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[112:113], v[240:241], v[76:79] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[114:115], v[242:243], v[76:79] \n"
" buffer_load_dwordx4 acc[248:251], %[v_os_b7], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[116:117], v[244:245], v[76:79] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[118:119], v[246:247], v[76:79] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[120:121], v[248:249], v[76:79] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[122:123], v[250:251], v[76:79] \n"
" buffer_load_dwordx4 acc[252:255], %[v_os_b7], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[124:125], v[252:253], v[76:79] \n"
_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[126:127], v[254:255], v[76:79] \n"
" s_mov_b64 exec, %[s_execflag_7] \n"
_UK_ATOMIC_ADD_ " %[v_os_o7], v17, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_add_u32 s60, 0x00000100, s80 \n"
" s_cmp_lt_u32 s60, %[s_loop_cnt] \n"
" s_cselect_b32 s56, %[s_tile_os_b], 0 \n"
" s_add_u32 s12, s56, s12 \n"
" s_addc_u32 s13, 0, s13 \n"
" s_cmp_ge_u32 s80, 0x00000100 \n"
" s_cselect_b32 s59, %[s_tile_os_o], s59 \n"
" s_add_u32 s8, s59, s8 \n"
" s_addc_u32 s9, 0, s9 \n"
" v_mul_f32 %[c0], %[scale_0], %[c0] \n"
" v_mul_f32 %[c1], %[scale_0], %[c1] \n"
" v_mul_f32 %[c2], %[scale_0], %[c2] \n"
" v_mul_f32 %[c3], %[scale_0], %[c3] \n"
" v_mul_f32 %[c4], %[scale_1], %[c4] \n"
" v_mul_f32 %[c5], %[scale_1], %[c5] \n"
" v_mul_f32 %[c6], %[scale_1], %[c6] \n"
" v_mul_f32 %[c7], %[scale_1], %[c7] \n"
" v_mul_f32 %[c8], %[scale_0], %[c8] \n"
" v_mul_f32 %[c9], %[scale_0], %[c9] \n"
" v_mul_f32 %[c10], %[scale_0], %[c10] \n"
" v_mul_f32 %[c11], %[scale_0], %[c11] \n"
" v_mul_f32 %[c12], %[scale_1], %[c12] \n"
" v_mul_f32 %[c13], %[scale_1], %[c13] \n"
" v_mul_f32 %[c14], %[scale_1], %[c14] \n"
" v_mul_f32 %[c15], %[scale_1], %[c15] \n"
_UK_PK_CVT_("%[c0]", "%[c1]", "%[c0]")
_UK_PK_CVT_("%[c2]", "%[c3]", "%[c1]")
_UK_PK_CVT_("%[c4]", "%[c5]", "%[c2]")
_UK_PK_CVT_("%[c6]", "%[c7]", "%[c3]")
_UK_PK_CVT_("%[c8]", "%[c9]", "%[c4]")
_UK_PK_CVT_("%[c10]", "%[c11]", "%[c5]")
_UK_PK_CVT_("%[c12]", "%[c13]", "%[c6]")
_UK_PK_CVT_("%[c14]", "%[c15]", "%[c7]")
" s_addk_i32 s80, 0x0080 \n"
" s_cmp_lt_i32 s80, %[s_loop_cnt] \n"
" s_cbranch_scc0 label_0EC1 \n"
" s_waitcnt vmcnt(30) & lgkmcnt(0) \n"
" s_barrier \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[128:129], v[128:129], 0 \n"
" ds_read_b32 v10, %[v_sfl_sld] offset:25344 \n"
" ds_read_b32 v11, %[v_sfl_sld] offset:25376 \n"
" ds_write_b64 v3, v[64:65] offset:16640 \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[130:131], v[130:131], v[80:83] \n"
" buffer_load_dwordx4 acc[0:3], %[v_os_b0], s[12:15], 0 offen \n"
" ds_write_b64 v3, v[66:67] offset:20992 \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[132:133], v[132:133], v[80:83] \n"
" ds_read_b32 v12, %[v_sfl_sld] offset:25408 \n"
" ds_read_b32 v13, %[v_sfl_sld] offset:25440 \n"
" ds_write_b64 v3, v[68:69] offset:18816 \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[134:135], v[134:135], v[80:83] \n"
" ds_write_b64 v3, v[70:71] offset:23168 \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[136:137], v[136:137], v[80:83] \n"
" ds_read_b32 v14, %[v_sfl_sld] offset:29696 \n"
" ds_read_b32 v15, %[v_sfl_sld] offset:29728 \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[138:139], v[138:139], v[80:83] \n"
" buffer_load_dwordx4 acc[4:7], %[v_os_b0], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[140:141], v[140:141], v[80:83] \n"
" ds_read_b32 v16, %[v_sfl_sld] offset:29760 \n"
" ds_read_b32 v17, %[v_sfl_sld] offset:29792 \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[142:143], v[142:143], v[80:83] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[128:129], v[192:193], 0 \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[130:131], v[194:195], v[84:87] \n"
" buffer_load_dwordx4 acc[8:11], %[v_os_b0], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[132:133], v[196:197], v[84:87] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[134:135], v[198:199], v[84:87] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[136:137], v[200:201], v[84:87] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[138:139], v[202:203], v[84:87] \n"
" buffer_load_dwordx4 acc[12:15], %[v_os_b0], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[140:141], v[204:205], v[84:87] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[142:143], v[206:207], v[84:87] \n"
" s_waitcnt lgkmcnt(0) \n"
" s_mov_b64 exec, %[s_execflag_0] \n"
_UK_ATOMIC_ADD_ " %[v_os_o0], v10, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[144:145], v[128:129], 0 \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[146:147], v[130:131], v[88:91] \n"
" buffer_load_dwordx4 acc[16:19], %[v_os_b1], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[148:149], v[132:133], v[88:91] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[150:151], v[134:135], v[88:91] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[152:153], v[136:137], v[88:91] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[154:155], v[138:139], v[88:91] \n"
" buffer_load_dwordx4 acc[20:23], %[v_os_b1], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[156:157], v[140:141], v[88:91] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[158:159], v[142:143], v[88:91] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[144:145], v[192:193], 0 \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[146:147], v[194:195], v[92:95] \n"
" buffer_load_dwordx4 acc[24:27], %[v_os_b1], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[148:149], v[196:197], v[92:95] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[150:151], v[198:199], v[92:95] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[152:153], v[200:201], v[92:95] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[154:155], v[202:203], v[92:95] \n"
" buffer_load_dwordx4 acc[28:31], %[v_os_b1], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[156:157], v[204:205], v[92:95] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[158:159], v[206:207], v[92:95] \n"
" s_mov_b64 exec, %[s_execflag_1] \n"
_UK_ATOMIC_ADD_ " %[v_os_o1], v11, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_waitcnt vmcnt(30) \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[160:161], v[144:145], v[80:83] \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[162:163], v[146:147], v[80:83] \n"
" buffer_load_dwordx4 acc[32:35], %[v_os_b2], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[164:165], v[148:149], v[80:83] \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[166:167], v[150:151], v[80:83] \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[168:169], v[152:153], v[80:83] \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[170:171], v[154:155], v[80:83] \n"
" buffer_load_dwordx4 acc[36:39], %[v_os_b2], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[172:173], v[156:157], v[80:83] \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[174:175], v[158:159], v[80:83] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[160:161], v[208:209], v[84:87] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[162:163], v[210:211], v[84:87] \n"
" buffer_load_dwordx4 acc[40:43], %[v_os_b2], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[164:165], v[212:213], v[84:87] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[166:167], v[214:215], v[84:87] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[168:169], v[216:217], v[84:87] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[170:171], v[218:219], v[84:87] \n"
" buffer_load_dwordx4 acc[44:47], %[v_os_b2], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[172:173], v[220:221], v[84:87] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[174:175], v[222:223], v[84:87] \n"
" s_mov_b64 exec, %[s_execflag_2] \n"
_UK_ATOMIC_ADD_ " %[v_os_o2], v12, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[176:177], v[144:145], v[88:91] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[178:179], v[146:147], v[88:91] \n"
" buffer_load_dwordx4 acc[48:51], %[v_os_b3], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[180:181], v[148:149], v[88:91] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[182:183], v[150:151], v[88:91] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[184:185], v[152:153], v[88:91] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[186:187], v[154:155], v[88:91] \n"
" buffer_load_dwordx4 acc[52:55], %[v_os_b3], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[188:189], v[156:157], v[88:91] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[190:191], v[158:159], v[88:91] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[176:177], v[208:209], v[92:95] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[178:179], v[210:211], v[92:95] \n"
" buffer_load_dwordx4 acc[56:59], %[v_os_b3], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[180:181], v[212:213], v[92:95] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[182:183], v[214:215], v[92:95] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[184:185], v[216:217], v[92:95] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[186:187], v[218:219], v[92:95] \n"
" buffer_load_dwordx4 acc[60:63], %[v_os_b3], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[188:189], v[220:221], v[92:95] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[190:191], v[222:223], v[92:95] \n"
" s_mov_b64 exec, %[s_execflag_3] \n"
_UK_ATOMIC_ADD_ " %[v_os_o3], v13, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_waitcnt vmcnt(30) \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[192:193], v[160:161], v[80:83] \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[194:195], v[162:163], v[80:83] \n"
" buffer_load_dwordx4 acc[64:67], %[v_os_b4], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[196:197], v[164:165], v[80:83] \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[198:199], v[166:167], v[80:83] \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[200:201], v[168:169], v[80:83] \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[202:203], v[170:171], v[80:83] \n"
" buffer_load_dwordx4 acc[68:71], %[v_os_b4], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[204:205], v[172:173], v[80:83] \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[206:207], v[174:175], v[80:83] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[192:193], v[224:225], v[84:87] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[194:195], v[226:227], v[84:87] \n"
" buffer_load_dwordx4 acc[72:75], %[v_os_b4], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[196:197], v[228:229], v[84:87] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[198:199], v[230:231], v[84:87] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[200:201], v[232:233], v[84:87] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[202:203], v[234:235], v[84:87] \n"
" buffer_load_dwordx4 acc[76:79], %[v_os_b4], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[204:205], v[236:237], v[84:87] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[206:207], v[238:239], v[84:87] \n"
" s_mov_b64 exec, %[s_execflag_4] \n"
_UK_ATOMIC_ADD_ " %[v_os_o4], v14, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[208:209], v[160:161], v[88:91] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[210:211], v[162:163], v[88:91] \n"
" buffer_load_dwordx4 acc[80:83], %[v_os_b5], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[212:213], v[164:165], v[88:91] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[214:215], v[166:167], v[88:91] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[216:217], v[168:169], v[88:91] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[218:219], v[170:171], v[88:91] \n"
" buffer_load_dwordx4 acc[84:87], %[v_os_b5], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[220:221], v[172:173], v[88:91] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[222:223], v[174:175], v[88:91] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[208:209], v[224:225], v[92:95] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[210:211], v[226:227], v[92:95] \n"
" buffer_load_dwordx4 acc[88:91], %[v_os_b5], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[212:213], v[228:229], v[92:95] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[214:215], v[230:231], v[92:95] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[216:217], v[232:233], v[92:95] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[218:219], v[234:235], v[92:95] \n"
" buffer_load_dwordx4 acc[92:95], %[v_os_b5], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[220:221], v[236:237], v[92:95] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[222:223], v[238:239], v[92:95] \n"
" s_mov_b64 exec, %[s_execflag_5] \n"
_UK_ATOMIC_ADD_ " %[v_os_o5], v15, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_waitcnt vmcnt(30) \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[224:225], v[176:177], v[80:83] \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[226:227], v[178:179], v[80:83] \n"
" buffer_load_dwordx4 acc[96:99], %[v_os_b6], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[228:229], v[180:181], v[80:83] \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[230:231], v[182:183], v[80:83] \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[232:233], v[184:185], v[80:83] \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[234:235], v[186:187], v[80:83] \n"
" buffer_load_dwordx4 acc[100:103], %[v_os_b6], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[236:237], v[188:189], v[80:83] \n"
_UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[238:239], v[190:191], v[80:83] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[224:225], v[240:241], v[84:87] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[226:227], v[242:243], v[84:87] \n"
" buffer_load_dwordx4 acc[104:107], %[v_os_b6], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[228:229], v[244:245], v[84:87] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[230:231], v[246:247], v[84:87] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[232:233], v[248:249], v[84:87] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[234:235], v[250:251], v[84:87] \n"
" buffer_load_dwordx4 acc[108:111], %[v_os_b6], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[236:237], v[252:253], v[84:87] \n"
_UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[238:239], v[254:255], v[84:87] \n"
" s_mov_b64 exec, %[s_execflag_6] \n"
_UK_ATOMIC_ADD_ " %[v_os_o6], v16, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[240:241], v[176:177], v[88:91] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[242:243], v[178:179], v[88:91] \n"
" buffer_load_dwordx4 acc[112:115], %[v_os_b7], s[12:15], 0 offen \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[244:245], v[180:181], v[88:91] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[246:247], v[182:183], v[88:91] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[248:249], v[184:185], v[88:91] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[250:251], v[186:187], v[88:91] \n"
" buffer_load_dwordx4 acc[116:119], %[v_os_b7], s[12:15], 0 offen offset:1024 \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[252:253], v[188:189], v[88:91] \n"
_UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[254:255], v[190:191], v[88:91] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[240:241], v[240:241], v[92:95] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[242:243], v[242:243], v[92:95] \n"
" buffer_load_dwordx4 acc[120:123], %[v_os_b7], s[12:15], 0 offen offset:2048 \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[244:245], v[244:245], v[92:95] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[246:247], v[246:247], v[92:95] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[248:249], v[248:249], v[92:95] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[250:251], v[250:251], v[92:95] \n"
" buffer_load_dwordx4 acc[124:127], %[v_os_b7], s[12:15], 0 offen offset:3072 \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[252:253], v[252:253], v[92:95] \n"
_UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[254:255], v[254:255], v[92:95] \n"
" s_mov_b64 exec, %[s_execflag_7] \n"
_UK_ATOMIC_ADD_ " %[v_os_o7], v17, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_add_u32 s60, 0x00000100, s80 \n"
" s_cmp_lt_u32 s60, %[s_loop_cnt] \n"
" s_cselect_b32 s56, s56, 0 \n"
" s_add_u32 s12, s56, s12 \n"
" s_addc_u32 s13, 0, s13 \n"
" s_cmp_ge_u32 s80, 0x00000100 \n"
" s_cselect_b32 s59, 0x00000100, s59 \n"
" s_add_u32 s8, s59, s8 \n"
" s_addc_u32 s9, 0, s9 \n"
" v_mul_f32 %[c16], %[scale_0], %[c16] \n"
" v_mul_f32 %[c17], %[scale_0], %[c17] \n"
" v_mul_f32 %[c18], %[scale_0], %[c18] \n"
" v_mul_f32 %[c19], %[scale_0], %[c19] \n"
" v_mul_f32 %[c20], %[scale_1], %[c20] \n"
" v_mul_f32 %[c21], %[scale_1], %[c21] \n"
" v_mul_f32 %[c22], %[scale_1], %[c22] \n"
" v_mul_f32 %[c23], %[scale_1], %[c23] \n"
" v_mul_f32 %[c24], %[scale_0], %[c24] \n"
" v_mul_f32 %[c25], %[scale_0], %[c25] \n"
" v_mul_f32 %[c26], %[scale_0], %[c26] \n"
" v_mul_f32 %[c27], %[scale_0], %[c27] \n"
" v_mul_f32 %[c28], %[scale_1], %[c28] \n"
" v_mul_f32 %[c29], %[scale_1], %[c29] \n"
" v_mul_f32 %[c30], %[scale_1], %[c30] \n"
" v_mul_f32 %[c31], %[scale_1], %[c31] \n"
_UK_PK_CVT_("%[c16]", "%[c17]", "%[c16]")
_UK_PK_CVT_("%[c18]", "%[c19]", "%[c17]")
_UK_PK_CVT_("%[c20]", "%[c21]", "%[c18]")
_UK_PK_CVT_("%[c22]", "%[c23]", "%[c19]")
_UK_PK_CVT_("%[c24]", "%[c25]", "%[c20]")
_UK_PK_CVT_("%[c26]", "%[c27]", "%[c21]")
_UK_PK_CVT_("%[c28]", "%[c29]", "%[c22]")
_UK_PK_CVT_("%[c30]", "%[c31]", "%[c23]")
" s_addk_i32 s80, 0x0080 \n"
" s_cmp_lt_i32 s80, %[s_loop_cnt] \n"
" s_cbranch_scc0 label_0EC1 \n"
" s_branch label_0AA6 \n"
" label_0EC1: \n"
" s_waitcnt lgkmcnt(0) \n"
" s_barrier \n"
" ds_read_b32 v10, %[v_sfl_sld] offset:16640 \n"
" ds_read_b32 v11, %[v_sfl_sld] offset:16672 \n"
" ds_read_b32 v12, %[v_sfl_sld] offset:16704 \n"
" ds_read_b32 v13, %[v_sfl_sld] offset:16736 \n"
" ds_read_b32 v14, %[v_sfl_sld] offset:20992 \n"
" ds_read_b32 v15, %[v_sfl_sld] offset:21024 \n"
" ds_read_b32 v16, %[v_sfl_sld] offset:21056 \n"
" ds_read_b32 v17, %[v_sfl_sld] offset:21088 \n"
" s_waitcnt lgkmcnt(0) \n"
" s_mov_b64 exec, %[s_execflag_0] \n"
_UK_ATOMIC_ADD_ " %[v_os_o0], v10, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_1] \n"
_UK_ATOMIC_ADD_ " %[v_os_o1], v11, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_2] \n"
_UK_ATOMIC_ADD_ " %[v_os_o2], v12, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_3] \n"
_UK_ATOMIC_ADD_ " %[v_os_o3], v13, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_4] \n"
_UK_ATOMIC_ADD_ " %[v_os_o4], v14, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_5] \n"
_UK_ATOMIC_ADD_ " %[v_os_o5], v15, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_6] \n"
_UK_ATOMIC_ADD_ " %[v_os_o6], v16, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_7] \n"
_UK_ATOMIC_ADD_ " %[v_os_o7], v17, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_add_u32 s8, s59, s8 \n"
" s_addc_u32 s9, 0, s9 \n"
" ds_write_b64 %[v_sfl_sst], [%[c16],%[c17]] offset:25344 \n"
" ds_write_b64 %[v_sfl_sst], [%[c18],%[c19]] offset:29696 \n"
" ds_write_b64 %[v_sfl_sst], [%[c20],%[c21]] offset:27520 \n"
" ds_write_b64 %[v_sfl_sst], [%[c22],%[c23]] offset:31872 \n"
" s_waitcnt lgkmcnt(0) \n"
" s_barrier \n"
" ds_read_b32 v10, %[v_sfl_sld] offset:25344 \n"
" ds_read_b32 v11, %[v_sfl_sld] offset:25376 \n"
" ds_read_b32 v12, %[v_sfl_sld] offset:25408 \n"
" ds_read_b32 v13, %[v_sfl_sld] offset:25440 \n"
" ds_read_b32 v14, %[v_sfl_sld] offset:29696 \n"
" ds_read_b32 v15, %[v_sfl_sld] offset:29728 \n"
" ds_read_b32 v16, %[v_sfl_sld] offset:29760 \n"
" ds_read_b32 v17, %[v_sfl_sld] offset:29792 \n"
" s_waitcnt lgkmcnt(0) \n"
" s_mov_b64 exec, %[s_execflag_0] \n"
_UK_ATOMIC_ADD_ " %[v_os_o0], v10, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_1] \n"
_UK_ATOMIC_ADD_ " %[v_os_o1], v11, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_2] \n"
_UK_ATOMIC_ADD_ " %[v_os_o2], v12, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_3] \n"
_UK_ATOMIC_ADD_ " %[v_os_o3], v13, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_4] \n"
_UK_ATOMIC_ADD_ " %[v_os_o4], v14, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_5] \n"
_UK_ATOMIC_ADD_ " %[v_os_o5], v15, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_6] \n"
_UK_ATOMIC_ADD_ " %[v_os_o6], v16, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_7] \n"
_UK_ATOMIC_ADD_ " %[v_os_o7], v17, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_mov_b32 s52, 0x07060302 ; v_perm\n"
" s_mov_b64 s[38:39], exec ; save current exec\n"
" s_mov_b32 s8, %[s_res_o0] \n"
" s_mov_b32 s9, %[s_res_o1] \n"
" s_mov_b32 s12, %[s_res_b0] \n"
" s_mov_b32 s13, %[s_res_b1] \n"
" s_mov_b32 s14, %[s_res_b2] \n"
" s_mov_b32 s15, %[s_res_b3] \n"
" s_mov_b32 s59, 0 \n"
" ds_read_b64 v[128:129], %[v_sld_y_os] offset:0 + %[sld_a_base] \n"
" ds_read_b64 v[130:131], %[v_sld_y_os] offset:128 + %[sld_a_base] \n"
" ds_read_b64 v[132:133], %[v_sld_y_os] offset:1024 + %[sld_a_base] \n"
" ds_read_b64 v[134:135], %[v_sld_y_os] offset:1152 + %[sld_a_base] \n"
" ds_read_b64 v[136:137], %[v_sld_y_os] offset:2048 + %[sld_a_base] \n"
" ds_read_b64 v[138:139], %[v_sld_y_os] offset:2176 + %[sld_a_base] \n"
" ds_read_b64 v[140:141], %[v_sld_y_os] offset:3072 + %[sld_a_base] \n"
" ds_read_b64 v[142:143], %[v_sld_y_os] offset:3200 + %[sld_a_base] \n"
" ds_read_b64 v[144:145], %[v_sld_y_os] offset:4096 + %[sld_a_base] \n"
" ds_read_b64 v[146:147], %[v_sld_y_os] offset:4224 + %[sld_a_base] \n"
" ds_read_b64 v[148:149], %[v_sld_y_os] offset:5120 + %[sld_a_base] \n"
" ds_read_b64 v[150:151], %[v_sld_y_os] offset:5248 + %[sld_a_base] \n"
" ds_read_b64 v[152:153], %[v_sld_y_os] offset:6144 + %[sld_a_base] \n"
" ds_read_b64 v[154:155], %[v_sld_y_os] offset:6272 + %[sld_a_base] \n"
" ds_read_b64 v[156:157], %[v_sld_y_os] offset:7168 + %[sld_a_base] \n"
" ds_read_b64 v[158:159], %[v_sld_y_os] offset:7296 + %[sld_a_base] \n"
" ds_read_b64 v[160:161], %[v_sld_y_os] offset:8192 + %[sld_a_base] \n"
" ds_read_b64 v[162:163], %[v_sld_y_os] offset:8320 + %[sld_a_base] \n"
" ds_read_b64 v[164:165], %[v_sld_y_os] offset:9216 + %[sld_a_base] \n"
" ds_read_b64 v[166:167], %[v_sld_y_os] offset:9344 + %[sld_a_base] \n"
" ds_read_b64 v[168:169], %[v_sld_y_os] offset:10240 + %[sld_a_base] \n"
" ds_read_b64 v[170:171], %[v_sld_y_os] offset:10368 + %[sld_a_base] \n"
" ds_read_b64 v[172:173], %[v_sld_y_os] offset:11264 + %[sld_a_base] \n"
" ds_read_b64 v[174:175], %[v_sld_y_os] offset:11392 + %[sld_a_base] \n"
" ds_read_b64 v[176:177], %[v_sld_y_os] offset:12288 + %[sld_a_base] \n"
" ds_read_b64 v[178:179], %[v_sld_y_os] offset:12416 + %[sld_a_base] \n"
" ds_read_b64 v[180:181], %[v_sld_y_os] offset:13312 + %[sld_a_base] \n"
" ds_read_b64 v[182:183], %[v_sld_y_os] offset:13440 + %[sld_a_base] \n"
" ds_read_b64 v[184:185], %[v_sld_y_os] offset:14336 + %[sld_a_base] \n"
" ds_read_b64 v[186:187], %[v_sld_y_os] offset:14464 + %[sld_a_base] \n"
" ds_read_b64 v[188:189], %[v_sld_y_os] offset:15360 + %[sld_a_base] \n"
" ds_read_b64 v[190:191], %[v_sld_y_os] offset:15488 + %[sld_a_base] \n"
" ds_read_b64 v[192:193], %[v_sld_y_os] offset:16384 + %[sld_a_base] \n"
" ds_read_b64 v[194:195], %[v_sld_y_os] offset:16512 + %[sld_a_base] \n"
" ds_read_b64 v[196:197], %[v_sld_y_os] offset:17408 + %[sld_a_base] \n"
" ds_read_b64 v[198:199], %[v_sld_y_os] offset:17536 + %[sld_a_base] \n"
" ds_read_b64 v[200:201], %[v_sld_y_os] offset:18432 + %[sld_a_base] \n"
" ds_read_b64 v[202:203], %[v_sld_y_os] offset:18560 + %[sld_a_base] \n"
" ds_read_b64 v[204:205], %[v_sld_y_os] offset:19456 + %[sld_a_base] \n"
" ds_read_b64 v[206:207], %[v_sld_y_os] offset:19584 + %[sld_a_base] \n"
" ds_read_b64 v[208:209], %[v_sld_y_os] offset:20480 + %[sld_a_base] \n"
" ds_read_b64 v[210:211], %[v_sld_y_os] offset:20608 + %[sld_a_base] \n"
" ds_read_b64 v[212:213], %[v_sld_y_os] offset:21504 + %[sld_a_base] \n"
" ds_read_b64 v[214:215], %[v_sld_y_os] offset:21632 + %[sld_a_base] \n"
" ds_read_b64 v[216:217], %[v_sld_y_os] offset:22528 + %[sld_a_base] \n"
" ds_read_b64 v[218:219], %[v_sld_y_os] offset:22656 + %[sld_a_base] \n"
" ds_read_b64 v[220:221], %[v_sld_y_os] offset:23552 + %[sld_a_base] \n"
" ds_read_b64 v[222:223], %[v_sld_y_os] offset:23680 + %[sld_a_base] \n"
" ds_read_b64 v[224:225], %[v_sld_y_os] offset:24576 + %[sld_a_base] \n"
" ds_read_b64 v[226:227], %[v_sld_y_os] offset:24704 + %[sld_a_base] \n"
" ds_read_b64 v[228:229], %[v_sld_y_os] offset:25600 + %[sld_a_base] \n"
" ds_read_b64 v[230:231], %[v_sld_y_os] offset:25728 + %[sld_a_base] \n"
" ds_read_b64 v[232:233], %[v_sld_y_os] offset:26624 + %[sld_a_base] \n"
" ds_read_b64 v[234:235], %[v_sld_y_os] offset:26752 + %[sld_a_base] \n"
" ds_read_b64 v[236:237], %[v_sld_y_os] offset:27648 + %[sld_a_base] \n"
" ds_read_b64 v[238:239], %[v_sld_y_os] offset:27776 + %[sld_a_base] \n"
" ds_read_b64 v[240:241], %[v_sld_y_os] offset:28672 + %[sld_a_base] \n"
" ds_read_b64 v[242:243], %[v_sld_y_os] offset:28800 + %[sld_a_base] \n"
" ds_read_b64 v[244:245], %[v_sld_y_os] offset:29696 + %[sld_a_base] \n"
" ds_read_b64 v[246:247], %[v_sld_y_os] offset:29824 + %[sld_a_base] \n"
" ds_read_b64 v[248:249], %[v_sld_y_os] offset:30720 + %[sld_a_base] \n"
" ds_read_b64 v[250:251], %[v_sld_y_os] offset:30848 + %[sld_a_base] \n"
" ds_read_b64 v[252:253], %[v_sld_y_os] offset:31744 + %[sld_a_base] \n"
" ds_read_b64 v[254:255], %[v_sld_y_os] offset:31872 + %[sld_a_base] \n"
" s_waitcnt 0 \n"
" buffer_load_dwordx4 acc[0:3], %[v_os_b0], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[4:7], %[v_os_b0], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[8:11], %[v_os_b0], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[12:15], %[v_os_b0], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[16:19], %[v_os_b1], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[20:23], %[v_os_b1], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[24:27], %[v_os_b1], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[28:31], %[v_os_b1], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[32:35], %[v_os_b2], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[36:39], %[v_os_b2], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[40:43], %[v_os_b2], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[44:47], %[v_os_b2], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[48:51], %[v_os_b3], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[52:55], %[v_os_b3], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[56:59], %[v_os_b3], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[60:63], %[v_os_b3], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[64:67], %[v_os_b4], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[68:71], %[v_os_b4], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[72:75], %[v_os_b4], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[76:79], %[v_os_b4], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[80:83], %[v_os_b5], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[84:87], %[v_os_b5], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[88:91], %[v_os_b5], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[92:95], %[v_os_b5], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[96:99], %[v_os_b6], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[100:103], %[v_os_b6], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[104:107], %[v_os_b6], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[108:111], %[v_os_b6], s[12:15], 0 offen offset:3072 \n"
" buffer_load_dwordx4 acc[112:115], %[v_os_b7], s[12:15], 0 offen \n"
" buffer_load_dwordx4 acc[116:119], %[v_os_b7], s[12:15], 0 offen offset:1024 \n"
" buffer_load_dwordx4 acc[120:123], %[v_os_b7], s[12:15], 0 offen offset:2048 \n"
" buffer_load_dwordx4 acc[124:127], %[v_os_b7], s[12:15], 0 offen offset:3072 \n"
" s_add_u32 s12, %[s_tile_os_b], s12 \n"
" s_addc_u32 s13, 0, s13 \n"
" v_mov_b32 v64, 0 \n"
" v_mov_b32 v80, 0 \n"
" v_mov_b32 v65, 0 \n"
" v_mov_b32 v81, 0 \n"
" v_mov_b32 v66, 0 \n"
" v_mov_b32 v82, 0 \n"
" v_mov_b32 v67, 0 \n"
" v_mov_b32 v83, 0 \n"
" v_mov_b32 v68, 0 \n"
" v_mov_b32 v84, 0 \n"
" v_mov_b32 v69, 0 \n"
" v_mov_b32 v85, 0 \n"
" v_mov_b32 v70, 0 \n"
" v_mov_b32 v86, 0 \n"
" v_mov_b32 v71, 0 \n"
" v_mov_b32 v87, 0 \n"
" ds_write_b64 %[v_sfl_sst], [%[c0],%[c1]] offset:16640 \n"
" ds_write_b64 %[v_sfl_sst], [%[c2],%[c3]] offset:20992 \n"
" ds_write_b64 %[v_sfl_sst], [%[c4],%[c5]] offset:18816 \n"
" ds_write_b64 %[v_sfl_sst], [%[c6],%[c7]] offset:23168 \n"
" s_mov_b32 s80, 0 \n"
" s_waitcnt vmcnt(24) \n"
"label_0AA6: \n"
" s_waitcnt vmcnt(30) & lgkmcnt(0) \n"
" s_barrier \n" _UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[0:1], v[128:129], 0 \n"
" ds_read_b32 v10, %[v_sfl_sld] offset:16640 \n"
" ds_read_b32 v11, %[v_sfl_sld] offset:16672 \n"
" ds_write_b64 %[v_sfl_sst], [%[c16],%[c17]] offset:25344 \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[2:3], v[130:131], v[64:67] \n"
" buffer_load_dwordx4 acc[128:131], %[v_os_b0], s[12:15], 0 offen \n"
" ds_write_b64 %[v_sfl_sst], [%[c18],%[c19]] offset:29696 \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[4:5], v[132:133], v[64:67] \n"
" ds_read_b32 v12, %[v_sfl_sld] offset:16704 \n"
" ds_read_b32 v13, %[v_sfl_sld] offset:16736 \n"
" ds_write_b64 %[v_sfl_sst], [%[c20],%[c21]] offset:27520 \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[6:7], v[134:135], v[64:67] \n"
" ds_write_b64 %[v_sfl_sst], [%[c22],%[c23]] offset:31872 \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[8:9], v[136:137], v[64:67] \n"
" ds_read_b32 v14, %[v_sfl_sld] offset:20992 \n"
" ds_read_b32 v15, %[v_sfl_sld] offset:21024 \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[10:11], v[138:139], v[64:67] \n"
" buffer_load_dwordx4 acc[132:135], %[v_os_b0], s[12:15], 0 offen offset:1024 \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[12:13], v[140:141], v[64:67] \n"
" ds_read_b32 v16, %[v_sfl_sld] offset:21056 \n"
" ds_read_b32 v17, %[v_sfl_sld] offset:21088 \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[14:15], v[142:143], v[64:67] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[0:1], v[192:193], 0 \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[2:3], v[194:195], v[68:71] \n"
" buffer_load_dwordx4 acc[136:139], %[v_os_b0], s[12:15], 0 offen offset:2048 \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[4:5], v[196:197], v[68:71] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[6:7], v[198:199], v[68:71] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[8:9], v[200:201], v[68:71] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[10:11], v[202:203], v[68:71] \n"
" buffer_load_dwordx4 acc[140:143], %[v_os_b0], s[12:15], 0 offen offset:3072 \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[12:13], v[204:205], v[68:71] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[14:15], v[206:207], v[68:71] \n"
" s_waitcnt lgkmcnt(0) \n"
" s_mov_b64 exec, %[s_execflag_0] \n" _UK_ATOMIC_ADD_ " %[v_os_o0], v10, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[16:17], v[128:129], 0 \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[18:19], v[130:131], v[72:75] \n"
" buffer_load_dwordx4 acc[144:147], %[v_os_b1], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[20:21], v[132:133], v[72:75] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[22:23], v[134:135], v[72:75] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[24:25], v[136:137], v[72:75] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[26:27], v[138:139], v[72:75] \n"
" buffer_load_dwordx4 acc[148:151], %[v_os_b1], s[12:15], 0 offen offset:1024 \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[28:29], v[140:141], v[72:75] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[30:31], v[142:143], v[72:75] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[16:17], v[192:193], 0 \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[18:19], v[194:195], v[76:79] \n"
" buffer_load_dwordx4 acc[152:155], %[v_os_b1], s[12:15], 0 offen offset:2048 \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[20:21], v[196:197], v[76:79] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[22:23], v[198:199], v[76:79] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[24:25], v[200:201], v[76:79] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[26:27], v[202:203], v[76:79] \n"
" buffer_load_dwordx4 acc[156:159], %[v_os_b1], s[12:15], 0 offen offset:3072 \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[28:29], v[204:205], v[76:79] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[30:31], v[206:207], v[76:79] \n"
" s_mov_b64 exec, %[s_execflag_1] \n" _UK_ATOMIC_ADD_ " %[v_os_o1], v11, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_waitcnt vmcnt(30) \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[32:33], v[144:145], v[64:67] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[34:35], v[146:147], v[64:67] \n"
" buffer_load_dwordx4 acc[160:163], %[v_os_b2], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[36:37], v[148:149], v[64:67] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[38:39], v[150:151], v[64:67] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[40:41], v[152:153], v[64:67] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[42:43], v[154:155], v[64:67] \n"
" buffer_load_dwordx4 acc[164:167], %[v_os_b2], s[12:15], 0 offen offset:1024 \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[44:45], v[156:157], v[64:67] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[46:47], v[158:159], v[64:67] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[32:33], v[208:209], v[68:71] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[34:35], v[210:211], v[68:71] \n"
" buffer_load_dwordx4 acc[168:171], %[v_os_b2], s[12:15], 0 offen offset:2048 \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[36:37], v[212:213], v[68:71] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[38:39], v[214:215], v[68:71] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[40:41], v[216:217], v[68:71] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[42:43], v[218:219], v[68:71] \n"
" buffer_load_dwordx4 acc[172:175], %[v_os_b2], s[12:15], 0 offen offset:3072 \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[44:45], v[220:221], v[68:71] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[46:47], v[222:223], v[68:71] \n"
" s_mov_b64 exec, %[s_execflag_2] \n" _UK_ATOMIC_ADD_ " %[v_os_o2], v12, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[48:49], v[144:145], v[72:75] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[50:51], v[146:147], v[72:75] \n"
" buffer_load_dwordx4 acc[176:179], %[v_os_b3], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[52:53], v[148:149], v[72:75] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[54:55], v[150:151], v[72:75] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[56:57], v[152:153], v[72:75] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[58:59], v[154:155], v[72:75] \n"
" buffer_load_dwordx4 acc[180:183], %[v_os_b3], s[12:15], 0 offen offset:1024 \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[60:61], v[156:157], v[72:75] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[62:63], v[158:159], v[72:75] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[48:49], v[208:209], v[76:79] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[50:51], v[210:211], v[76:79] \n"
" buffer_load_dwordx4 acc[184:187], %[v_os_b3], s[12:15], 0 offen offset:2048 \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[52:53], v[212:213], v[76:79] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[54:55], v[214:215], v[76:79] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[56:57], v[216:217], v[76:79] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[58:59], v[218:219], v[76:79] \n"
" buffer_load_dwordx4 acc[188:191], %[v_os_b3], s[12:15], 0 offen offset:3072 \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[60:61], v[220:221], v[76:79] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[62:63], v[222:223], v[76:79] \n"
" s_mov_b64 exec, %[s_execflag_3] \n" _UK_ATOMIC_ADD_ " %[v_os_o3], v13, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_waitcnt vmcnt(30) \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[64:65], v[160:161], v[64:67] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[66:67], v[162:163], v[64:67] \n"
" buffer_load_dwordx4 acc[192:195], %[v_os_b4], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[68:69], v[164:165], v[64:67] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[70:71], v[166:167], v[64:67] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[72:73], v[168:169], v[64:67] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[74:75], v[170:171], v[64:67] \n"
" buffer_load_dwordx4 acc[196:199], %[v_os_b4], s[12:15], 0 offen offset:1024 \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[76:77], v[172:173], v[64:67] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[78:79], v[174:175], v[64:67] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[64:65], v[224:225], v[68:71] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[66:67], v[226:227], v[68:71] \n"
" buffer_load_dwordx4 acc[200:203], %[v_os_b4], s[12:15], 0 offen offset:2048 \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[68:69], v[228:229], v[68:71] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[70:71], v[230:231], v[68:71] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[72:73], v[232:233], v[68:71] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[74:75], v[234:235], v[68:71] \n"
" buffer_load_dwordx4 acc[204:207], %[v_os_b4], s[12:15], 0 offen offset:3072 \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[76:77], v[236:237], v[68:71] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[78:79], v[238:239], v[68:71] \n"
" s_mov_b64 exec, %[s_execflag_4] \n" _UK_ATOMIC_ADD_ " %[v_os_o4], v14, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[80:81], v[160:161], v[72:75] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[82:83], v[162:163], v[72:75] \n"
" buffer_load_dwordx4 acc[208:211], %[v_os_b5], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[84:85], v[164:165], v[72:75] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[86:87], v[166:167], v[72:75] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[88:89], v[168:169], v[72:75] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[90:91], v[170:171], v[72:75] \n"
" buffer_load_dwordx4 acc[212:215], %[v_os_b5], s[12:15], 0 offen offset:1024 \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[92:93], v[172:173], v[72:75] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[94:95], v[174:175], v[72:75] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[80:81], v[224:225], v[76:79] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[82:83], v[226:227], v[76:79] \n"
" buffer_load_dwordx4 acc[216:219], %[v_os_b5], s[12:15], 0 offen offset:2048 \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[84:85], v[228:229], v[76:79] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[86:87], v[230:231], v[76:79] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[88:89], v[232:233], v[76:79] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[90:91], v[234:235], v[76:79] \n"
" buffer_load_dwordx4 acc[220:223], %[v_os_b5], s[12:15], 0 offen offset:3072 \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[92:93], v[236:237], v[76:79] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[94:95], v[238:239], v[76:79] \n"
" s_mov_b64 exec, %[s_execflag_5] \n" _UK_ATOMIC_ADD_ " %[v_os_o5], v15, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_waitcnt vmcnt(30) \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[96:97], v[176:177], v[64:67] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[98:99], v[178:179], v[64:67] \n"
" buffer_load_dwordx4 acc[224:227], %[v_os_b6], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[100:101], v[180:181], v[64:67] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[102:103], v[182:183], v[64:67] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[104:105], v[184:185], v[64:67] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[106:107], v[186:187], v[64:67] \n"
" buffer_load_dwordx4 acc[228:231], %[v_os_b6], s[12:15], 0 offen offset:1024 \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[108:109], v[188:189], v[64:67] \n" _UK_MFMA_
" [%[c0], %[c1], %[c2], %[c3]], acc[110:111], v[190:191], v[64:67] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[96:97], v[240:241], v[68:71] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[98:99], v[242:243], v[68:71] \n"
" buffer_load_dwordx4 acc[232:235], %[v_os_b6], s[12:15], 0 offen offset:2048 \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[100:101], v[244:245], v[68:71] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[102:103], v[246:247], v[68:71] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[104:105], v[248:249], v[68:71] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[106:107], v[250:251], v[68:71] \n"
" buffer_load_dwordx4 acc[236:239], %[v_os_b6], s[12:15], 0 offen offset:3072 \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[108:109], v[252:253], v[68:71] \n" _UK_MFMA_
" [%[c4], %[c5], %[c6], %[c7]], acc[110:111], v[254:255], v[68:71] \n"
" s_mov_b64 exec, %[s_execflag_6] \n" _UK_ATOMIC_ADD_ " %[v_os_o6], v16, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[112:113], v[176:177], v[72:75] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[114:115], v[178:179], v[72:75] \n"
" buffer_load_dwordx4 acc[240:243], %[v_os_b7], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[116:117], v[180:181], v[72:75] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[118:119], v[182:183], v[72:75] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[120:121], v[184:185], v[72:75] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[122:123], v[186:187], v[72:75] \n"
" buffer_load_dwordx4 acc[244:247], %[v_os_b7], s[12:15], 0 offen offset:1024 \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[124:125], v[188:189], v[72:75] \n" _UK_MFMA_
" [%[c8], %[c9], %[c10], %[c11]], acc[126:127], v[190:191], v[72:75] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[112:113], v[240:241], v[76:79] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[114:115], v[242:243], v[76:79] \n"
" buffer_load_dwordx4 acc[248:251], %[v_os_b7], s[12:15], 0 offen offset:2048 \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[116:117], v[244:245], v[76:79] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[118:119], v[246:247], v[76:79] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[120:121], v[248:249], v[76:79] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[122:123], v[250:251], v[76:79] \n"
" buffer_load_dwordx4 acc[252:255], %[v_os_b7], s[12:15], 0 offen offset:3072 \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[124:125], v[252:253], v[76:79] \n" _UK_MFMA_
" [%[c12], %[c13], %[c14], %[c15]], acc[126:127], v[254:255], v[76:79] \n"
" s_mov_b64 exec, %[s_execflag_7] \n" _UK_ATOMIC_ADD_ " %[v_os_o7], v17, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_add_u32 s60, 0x00000100, s80 \n"
" s_cmp_lt_u32 s60, %[s_loop_cnt] \n"
" s_cselect_b32 s56, %[s_tile_os_b], 0 \n"
" s_add_u32 s12, s56, s12 \n"
" s_addc_u32 s13, 0, s13 \n"
" s_cmp_ge_u32 s80, 0x00000100 \n"
" s_cselect_b32 s59, %[s_tile_os_o], s59 \n"
" s_add_u32 s8, s59, s8 \n"
" s_addc_u32 s9, 0, s9 \n"
" v_mul_f32 %[c0], %[scale_0], %[c0] \n"
" v_mul_f32 %[c1], %[scale_0], %[c1] \n"
" v_mul_f32 %[c2], %[scale_0], %[c2] \n"
" v_mul_f32 %[c3], %[scale_0], %[c3] \n"
" v_mul_f32 %[c4], %[scale_1], %[c4] \n"
" v_mul_f32 %[c5], %[scale_1], %[c5] \n"
" v_mul_f32 %[c6], %[scale_1], %[c6] \n"
" v_mul_f32 %[c7], %[scale_1], %[c7] \n"
" v_mul_f32 %[c8], %[scale_0], %[c8] \n"
" v_mul_f32 %[c9], %[scale_0], %[c9] \n"
" v_mul_f32 %[c10], %[scale_0], %[c10] \n"
" v_mul_f32 %[c11], %[scale_0], %[c11] \n"
" v_mul_f32 %[c12], %[scale_1], %[c12] \n"
" v_mul_f32 %[c13], %[scale_1], %[c13] \n"
" v_mul_f32 %[c14], %[scale_1], %[c14] \n"
" v_mul_f32 %[c15], %[scale_1], %[c15] \n" _UK_PK_CVT_("%[c0]", "%[c1]", "%[c0]") _UK_PK_CVT_(
"%[c2]",
"%[c3]",
"%[c1]") _UK_PK_CVT_("%[c4]",
"%[c5]",
"%[c2]") _UK_PK_CVT_("%[c6]",
"%[c7]",
"%[c3]") _UK_PK_CVT_("%[c8]",
"%[c9]",
"%[c4]") _UK_PK_CVT_("%["
"c10]",
"%["
"c11]",
"%[c5]")
_UK_PK_CVT_("%[c12]", "%[c13]", "%[c6]") _UK_PK_CVT_(
"%[c14]",
"%[c15]",
"%[c7]") " s_addk_i32 s80, 0x0080 \n"
" s_cmp_lt_i32 s80, %[s_loop_cnt] \n"
" s_cbranch_scc0 label_0EC1 \n"
" s_waitcnt vmcnt(30) & lgkmcnt(0) \n"
" s_barrier \n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[128:129], v[128:129], 0 \n"
" ds_read_b32 v10, %[v_sfl_sld] offset:25344 \n"
" ds_read_b32 v11, %[v_sfl_sld] offset:25376 \n"
" ds_write_b64 v3, v[64:65] offset:16640 \n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[130:131], v[130:131], v[80:83] \n"
" buffer_load_dwordx4 acc[0:3], %[v_os_b0], s[12:15], 0 offen \n"
" ds_write_b64 v3, v[66:67] offset:20992 \n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[132:133], v[132:133], v[80:83] \n"
" ds_read_b32 v12, %[v_sfl_sld] offset:25408 \n"
" ds_read_b32 v13, %[v_sfl_sld] offset:25440 \n"
" ds_write_b64 v3, v[68:69] offset:18816 \n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[134:135], v[134:135], v[80:83] \n"
" ds_write_b64 v3, v[70:71] offset:23168 \n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[136:137], v[136:137], v[80:83] \n"
" ds_read_b32 v14, %[v_sfl_sld] offset:29696 \n"
" ds_read_b32 v15, %[v_sfl_sld] offset:29728 \n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[138:139], v[138:139], v[80:83] \n"
" buffer_load_dwordx4 acc[4:7], %[v_os_b0], s[12:15], 0 offen offset:1024 "
"\n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[140:141], v[140:141], v[80:83] \n"
" ds_read_b32 v16, %[v_sfl_sld] offset:29760 \n"
" ds_read_b32 v17, %[v_sfl_sld] offset:29792 \n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[142:143], v[142:143], v[80:83] "
"\n" _UK_MFMA_
" [%[c20], %[c21], %[c22], %[c23]], acc[128:129], v[192:193], 0 \n" _UK_MFMA_
" [%[c20], %[c21], %[c22], %[c23]], acc[130:131], v[194:195], v[84:87] \n"
" buffer_load_dwordx4 acc[8:11], %[v_os_b0], s[12:15], 0 offen offset:2048 "
"\n" _UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[132:133], v[196:197], "
"v[84:87] \n" _UK_MFMA_
" [%[c20], %[c21], %[c22], %[c23]], acc[134:135], v[198:199], v[84:87] "
"\n" _UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[136:137], v[200:201], "
"v[84:87] \n" _UK_MFMA_
" [%[c20], %[c21], %[c22], %[c23]], acc[138:139], v[202:203], v[84:87] \n"
" buffer_load_dwordx4 acc[12:15], %[v_os_b0], s[12:15], 0 offen offset:3072 "
"\n" _UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[140:141], v[204:205], "
"v[84:87] \n" _UK_MFMA_
" [%[c20], %[c21], %[c22], %[c23]], acc[142:143], v[206:207], v[84:87] \n"
" s_waitcnt lgkmcnt(0) \n"
" s_mov_b64 exec, %[s_execflag_0] \n" _UK_ATOMIC_ADD_
" %[v_os_o0], v10, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[144:145], v[128:129], 0 \n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[146:147], v[130:131], v[88:91] \n"
" buffer_load_dwordx4 acc[16:19], %[v_os_b1], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[148:149], v[132:133], v[88:91] "
"\n" _UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[150:151], v[134:135], "
"v[88:91] \n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[152:153], v[136:137], v[88:91] "
"\n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[154:155], v[138:139], v[88:91] \n"
" buffer_load_dwordx4 acc[20:23], %[v_os_b1], s[12:15], 0 offen offset:1024 "
"\n" _UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[156:157], v[140:141], "
"v[88:91] \n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[158:159], v[142:143], v[88:91] "
"\n" _UK_MFMA_
" [%[c28], %[c29], %[c30], %[c31]], acc[144:145], v[192:193], 0 \n" _UK_MFMA_
" [%[c28], %[c29], %[c30], %[c31]], acc[146:147], v[194:195], v[92:95] \n"
" buffer_load_dwordx4 acc[24:27], %[v_os_b1], s[12:15], 0 offen offset:2048 "
"\n" _UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[148:149], v[196:197], "
"v[92:95] \n" _UK_MFMA_
" [%[c28], %[c29], %[c30], %[c31]], acc[150:151], v[198:199], v[92:95] "
"\n" _UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[152:153], v[200:201], "
"v[92:95] \n" _UK_MFMA_
" [%[c28], %[c29], %[c30], %[c31]], acc[154:155], v[202:203], v[92:95] \n"
" buffer_load_dwordx4 acc[28:31], %[v_os_b1], s[12:15], 0 offen offset:3072 "
"\n" _UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[156:157], v[204:205], "
"v[92:95] \n" _UK_MFMA_
" [%[c28], %[c29], %[c30], %[c31]], acc[158:159], v[206:207], v[92:95] \n"
" s_mov_b64 exec, %[s_execflag_1] \n" _UK_ATOMIC_ADD_
" %[v_os_o1], v11, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_waitcnt vmcnt(30) \n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[160:161], v[144:145], v[80:83] "
"\n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[162:163], v[146:147], v[80:83] \n"
" buffer_load_dwordx4 acc[32:35], %[v_os_b2], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[164:165], v[148:149], v[80:83] "
"\n" _UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[166:167], v[150:151], "
"v[80:83] \n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[168:169], v[152:153], v[80:83] "
"\n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[170:171], v[154:155], v[80:83] \n"
" buffer_load_dwordx4 acc[36:39], %[v_os_b2], s[12:15], 0 offen offset:1024 "
"\n" _UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[172:173], v[156:157], "
"v[80:83] \n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[174:175], v[158:159], v[80:83] "
"\n" _UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[160:161], v[208:209], "
"v[84:87] \n" _UK_MFMA_
" [%[c20], %[c21], %[c22], %[c23]], acc[162:163], v[210:211], v[84:87] \n"
" buffer_load_dwordx4 acc[40:43], %[v_os_b2], s[12:15], 0 offen offset:2048 "
"\n" _UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[164:165], v[212:213], "
"v[84:87] \n" _UK_MFMA_
" [%[c20], %[c21], %[c22], %[c23]], acc[166:167], v[214:215], v[84:87] "
"\n" _UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[168:169], v[216:217], "
"v[84:87] \n" _UK_MFMA_
" [%[c20], %[c21], %[c22], %[c23]], acc[170:171], v[218:219], v[84:87] \n"
" buffer_load_dwordx4 acc[44:47], %[v_os_b2], s[12:15], 0 offen offset:3072 "
"\n" _UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[172:173], v[220:221], "
"v[84:87] \n" _UK_MFMA_
" [%[c20], %[c21], %[c22], %[c23]], acc[174:175], v[222:223], v[84:87] \n"
" s_mov_b64 exec, %[s_execflag_2] \n" _UK_ATOMIC_ADD_
" %[v_os_o2], v12, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[176:177], v[144:145], v[88:91] "
"\n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[178:179], v[146:147], v[88:91] \n"
" buffer_load_dwordx4 acc[48:51], %[v_os_b3], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[180:181], v[148:149], v[88:91] "
"\n" _UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[182:183], v[150:151], "
"v[88:91] \n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[184:185], v[152:153], v[88:91] "
"\n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[186:187], v[154:155], v[88:91] \n"
" buffer_load_dwordx4 acc[52:55], %[v_os_b3], s[12:15], 0 offen offset:1024 "
"\n" _UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[188:189], v[156:157], "
"v[88:91] \n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[190:191], v[158:159], v[88:91] "
"\n" _UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[176:177], v[208:209], "
"v[92:95] \n" _UK_MFMA_
" [%[c28], %[c29], %[c30], %[c31]], acc[178:179], v[210:211], v[92:95] \n"
" buffer_load_dwordx4 acc[56:59], %[v_os_b3], s[12:15], 0 offen offset:2048 "
"\n" _UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[180:181], v[212:213], "
"v[92:95] \n" _UK_MFMA_
" [%[c28], %[c29], %[c30], %[c31]], acc[182:183], v[214:215], v[92:95] "
"\n" _UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[184:185], v[216:217], "
"v[92:95] \n" _UK_MFMA_
" [%[c28], %[c29], %[c30], %[c31]], acc[186:187], v[218:219], v[92:95] \n"
" buffer_load_dwordx4 acc[60:63], %[v_os_b3], s[12:15], 0 offen offset:3072 "
"\n" _UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[188:189], v[220:221], "
"v[92:95] \n" _UK_MFMA_
" [%[c28], %[c29], %[c30], %[c31]], acc[190:191], v[222:223], v[92:95] \n"
" s_mov_b64 exec, %[s_execflag_3] \n" _UK_ATOMIC_ADD_
" %[v_os_o3], v13, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_waitcnt vmcnt(30) \n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[192:193], v[160:161], v[80:83] "
"\n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[194:195], v[162:163], v[80:83] \n"
" buffer_load_dwordx4 acc[64:67], %[v_os_b4], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[196:197], v[164:165], v[80:83] "
"\n" _UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[198:199], v[166:167], "
"v[80:83] \n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[200:201], v[168:169], v[80:83] "
"\n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[202:203], v[170:171], v[80:83] \n"
" buffer_load_dwordx4 acc[68:71], %[v_os_b4], s[12:15], 0 offen offset:1024 "
"\n" _UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[204:205], v[172:173], "
"v[80:83] \n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[206:207], v[174:175], v[80:83] "
"\n" _UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[192:193], v[224:225], "
"v[84:87] \n" _UK_MFMA_
" [%[c20], %[c21], %[c22], %[c23]], acc[194:195], v[226:227], v[84:87] \n"
" buffer_load_dwordx4 acc[72:75], %[v_os_b4], s[12:15], 0 offen offset:2048 "
"\n" _UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[196:197], v[228:229], "
"v[84:87] \n" _UK_MFMA_
" [%[c20], %[c21], %[c22], %[c23]], acc[198:199], v[230:231], v[84:87] "
"\n" _UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[200:201], v[232:233], "
"v[84:87] \n" _UK_MFMA_
" [%[c20], %[c21], %[c22], %[c23]], acc[202:203], v[234:235], v[84:87] \n"
" buffer_load_dwordx4 acc[76:79], %[v_os_b4], s[12:15], 0 offen offset:3072 "
"\n" _UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[204:205], v[236:237], "
"v[84:87] \n" _UK_MFMA_
" [%[c20], %[c21], %[c22], %[c23]], acc[206:207], v[238:239], v[84:87] \n"
" s_mov_b64 exec, %[s_execflag_4] \n" _UK_ATOMIC_ADD_
" %[v_os_o4], v14, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[208:209], v[160:161], v[88:91] "
"\n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[210:211], v[162:163], v[88:91] \n"
" buffer_load_dwordx4 acc[80:83], %[v_os_b5], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[212:213], v[164:165], v[88:91] "
"\n" _UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[214:215], v[166:167], "
"v[88:91] \n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[216:217], v[168:169], v[88:91] "
"\n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[218:219], v[170:171], v[88:91] \n"
" buffer_load_dwordx4 acc[84:87], %[v_os_b5], s[12:15], 0 offen offset:1024 "
"\n" _UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[220:221], v[172:173], "
"v[88:91] \n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[222:223], v[174:175], v[88:91] "
"\n" _UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[208:209], v[224:225], "
"v[92:95] \n" _UK_MFMA_
" [%[c28], %[c29], %[c30], %[c31]], acc[210:211], v[226:227], v[92:95] \n"
" buffer_load_dwordx4 acc[88:91], %[v_os_b5], s[12:15], 0 offen offset:2048 "
"\n" _UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[212:213], v[228:229], "
"v[92:95] \n" _UK_MFMA_
" [%[c28], %[c29], %[c30], %[c31]], acc[214:215], v[230:231], v[92:95] "
"\n" _UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[216:217], v[232:233], "
"v[92:95] \n" _UK_MFMA_
" [%[c28], %[c29], %[c30], %[c31]], acc[218:219], v[234:235], v[92:95] \n"
" buffer_load_dwordx4 acc[92:95], %[v_os_b5], s[12:15], 0 offen offset:3072 "
"\n" _UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[220:221], v[236:237], "
"v[92:95] \n" _UK_MFMA_
" [%[c28], %[c29], %[c30], %[c31]], acc[222:223], v[238:239], v[92:95] \n"
" s_mov_b64 exec, %[s_execflag_5] \n" _UK_ATOMIC_ADD_
" %[v_os_o5], v15, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_waitcnt vmcnt(30) \n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[224:225], v[176:177], v[80:83] "
"\n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[226:227], v[178:179], v[80:83] \n"
" buffer_load_dwordx4 acc[96:99], %[v_os_b6], s[12:15], 0 offen \n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[228:229], v[180:181], v[80:83] "
"\n" _UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[230:231], v[182:183], "
"v[80:83] \n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[232:233], v[184:185], v[80:83] "
"\n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[234:235], v[186:187], v[80:83] \n"
" buffer_load_dwordx4 acc[100:103], %[v_os_b6], s[12:15], 0 offen "
"offset:1024 \n" _UK_MFMA_ " [%[c16], %[c17], %[c18], %[c19]], acc[236:237], "
"v[188:189], v[80:83] \n" _UK_MFMA_
" [%[c16], %[c17], %[c18], %[c19]], acc[238:239], v[190:191], v[80:83] "
"\n" _UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[224:225], v[240:241], "
"v[84:87] \n" _UK_MFMA_
" [%[c20], %[c21], %[c22], %[c23]], acc[226:227], v[242:243], v[84:87] \n"
" buffer_load_dwordx4 acc[104:107], %[v_os_b6], s[12:15], 0 offen "
"offset:2048 \n" _UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[228:229], "
"v[244:245], v[84:87] \n" _UK_MFMA_
" [%[c20], %[c21], %[c22], %[c23]], acc[230:231], v[246:247], v[84:87] "
"\n" _UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[232:233], v[248:249], "
"v[84:87] \n" _UK_MFMA_
" [%[c20], %[c21], %[c22], %[c23]], acc[234:235], v[250:251], v[84:87] \n"
" buffer_load_dwordx4 acc[108:111], %[v_os_b6], s[12:15], 0 offen "
"offset:3072 \n" _UK_MFMA_ " [%[c20], %[c21], %[c22], %[c23]], acc[236:237], "
"v[252:253], v[84:87] \n" _UK_MFMA_
" [%[c20], %[c21], %[c22], %[c23]], acc[238:239], v[254:255], v[84:87] \n"
" s_mov_b64 exec, %[s_execflag_6] \n" _UK_ATOMIC_ADD_
" %[v_os_o6], v16, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[240:241], v[176:177], v[88:91] "
"\n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[242:243], v[178:179], v[88:91] \n"
" buffer_load_dwordx4 acc[112:115], %[v_os_b7], s[12:15], 0 offen "
"\n" _UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[244:245], v[180:181], "
"v[88:91] \n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[246:247], v[182:183], v[88:91] "
"\n" _UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[248:249], v[184:185], "
"v[88:91] \n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[250:251], v[186:187], v[88:91] \n"
" buffer_load_dwordx4 acc[116:119], %[v_os_b7], s[12:15], 0 offen "
"offset:1024 \n" _UK_MFMA_ " [%[c24], %[c25], %[c26], %[c27]], acc[252:253], "
"v[188:189], v[88:91] \n" _UK_MFMA_
" [%[c24], %[c25], %[c26], %[c27]], acc[254:255], v[190:191], v[88:91] "
"\n" _UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[240:241], v[240:241], "
"v[92:95] \n" _UK_MFMA_
" [%[c28], %[c29], %[c30], %[c31]], acc[242:243], v[242:243], v[92:95] \n"
" buffer_load_dwordx4 acc[120:123], %[v_os_b7], s[12:15], 0 offen "
"offset:2048 \n" _UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[244:245], "
"v[244:245], v[92:95] \n" _UK_MFMA_
" [%[c28], %[c29], %[c30], %[c31]], acc[246:247], v[246:247], v[92:95] "
"\n" _UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[248:249], v[248:249], "
"v[92:95] \n" _UK_MFMA_
" [%[c28], %[c29], %[c30], %[c31]], acc[250:251], v[250:251], v[92:95] \n"
" buffer_load_dwordx4 acc[124:127], %[v_os_b7], s[12:15], 0 offen "
"offset:3072 \n" _UK_MFMA_ " [%[c28], %[c29], %[c30], %[c31]], acc[252:253], "
"v[252:253], v[92:95] \n" _UK_MFMA_
" [%[c28], %[c29], %[c30], %[c31]], acc[254:255], v[254:255], v[92:95] \n"
" s_mov_b64 exec, %[s_execflag_7] \n" _UK_ATOMIC_ADD_
" %[v_os_o7], v17, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
" s_add_u32 s60, 0x00000100, s80 \n"
" s_cmp_lt_u32 s60, %[s_loop_cnt] \n"
" s_cselect_b32 s56, s56, 0 \n"
" s_add_u32 s12, s56, s12 \n"
" s_addc_u32 s13, 0, s13 \n"
" s_cmp_ge_u32 s80, 0x00000100 \n"
" s_cselect_b32 s59, 0x00000100, s59 \n"
" s_add_u32 s8, s59, s8 \n"
" s_addc_u32 s9, 0, s9 \n"
" v_mul_f32 %[c16], %[scale_0], %[c16] \n"
" v_mul_f32 %[c17], %[scale_0], %[c17] \n"
" v_mul_f32 %[c18], %[scale_0], %[c18] \n"
" v_mul_f32 %[c19], %[scale_0], %[c19] \n"
" v_mul_f32 %[c20], %[scale_1], %[c20] \n"
" v_mul_f32 %[c21], %[scale_1], %[c21] \n"
" v_mul_f32 %[c22], %[scale_1], %[c22] \n"
" v_mul_f32 %[c23], %[scale_1], %[c23] \n"
" v_mul_f32 %[c24], %[scale_0], %[c24] \n"
" v_mul_f32 %[c25], %[scale_0], %[c25] \n"
" v_mul_f32 %[c26], %[scale_0], %[c26] \n"
" v_mul_f32 %[c27], %[scale_0], %[c27] \n"
" v_mul_f32 %[c28], %[scale_1], %[c28] \n"
" v_mul_f32 %[c29], %[scale_1], %[c29] \n"
" v_mul_f32 %[c30], %[scale_1], %[c30] \n"
" v_mul_f32 %[c31], %[scale_1], %[c31] \n" _UK_PK_CVT_(
"%[c16]", "%[c17]", "%[c16]") _UK_PK_CVT_("%[c18]", "%[c19]", "%[c17]")
_UK_PK_CVT_("%[c20]", "%[c21]", "%[c18]") _UK_PK_CVT_(
"%[c22]", "%[c23]", "%[c19]") _UK_PK_CVT_("%[c24]", "%[c25]", "%[c20]")
_UK_PK_CVT_("%[c26]", "%[c27]", "%[c21]")
_UK_PK_CVT_("%[c28]", "%[c29]", "%[c22]") _UK_PK_CVT_(
"%[c30]",
"%[c31]",
"%[c23]") " s_addk_i32 s80, 0x0080 \n"
" s_cmp_lt_i32 s80, %[s_loop_cnt] \n"
" s_cbranch_scc0 label_0EC1 \n"
" s_branch label_0AA6 \n"
" label_0EC1: \n"
" s_waitcnt lgkmcnt(0) \n"
" s_barrier \n"
" ds_read_b32 v10, %[v_sfl_sld] offset:16640 \n"
" ds_read_b32 v11, %[v_sfl_sld] offset:16672 \n"
" ds_read_b32 v12, %[v_sfl_sld] offset:16704 \n"
" ds_read_b32 v13, %[v_sfl_sld] offset:16736 \n"
" ds_read_b32 v14, %[v_sfl_sld] offset:20992 \n"
" ds_read_b32 v15, %[v_sfl_sld] offset:21024 \n"
" ds_read_b32 v16, %[v_sfl_sld] offset:21056 \n"
" ds_read_b32 v17, %[v_sfl_sld] offset:21088 \n"
" s_waitcnt lgkmcnt(0) \n"
" s_mov_b64 exec, %[s_execflag_0] \n" _UK_ATOMIC_ADD_
" %[v_os_o0], v10, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_1] \n" _UK_ATOMIC_ADD_
" %[v_os_o1], v11, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_2] \n" _UK_ATOMIC_ADD_
" %[v_os_o2], v12, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_3] \n" _UK_ATOMIC_ADD_
" %[v_os_o3], v13, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_4] \n" _UK_ATOMIC_ADD_
" %[v_os_o4], v14, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_5] \n" _UK_ATOMIC_ADD_
" %[v_os_o5], v15, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_6] \n" _UK_ATOMIC_ADD_
" %[v_os_o6], v16, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_7] \n" _UK_ATOMIC_ADD_
" %[v_os_o7], v17, s[8:9] \n"
" s_mov_b64 exec, s[38:39] "
" \n"
" s_add_u32 s8, s59, s8 \n"
" s_addc_u32 s9, 0, s9 \n"
" ds_write_b64 %[v_sfl_sst], [%[c16],%[c17]] "
"offset:25344 \n"
" ds_write_b64 %[v_sfl_sst], [%[c18],%[c19]] "
"offset:29696 \n"
" ds_write_b64 %[v_sfl_sst], [%[c20],%[c21]] "
"offset:27520 \n"
" ds_write_b64 %[v_sfl_sst], [%[c22],%[c23]] "
"offset:31872 \n"
" s_waitcnt lgkmcnt(0) \n"
" s_barrier \n"
" ds_read_b32 v10, %[v_sfl_sld] offset:25344 \n"
" ds_read_b32 v11, %[v_sfl_sld] offset:25376 \n"
" ds_read_b32 v12, %[v_sfl_sld] offset:25408 \n"
" ds_read_b32 v13, %[v_sfl_sld] offset:25440 \n"
" ds_read_b32 v14, %[v_sfl_sld] offset:29696 \n"
" ds_read_b32 v15, %[v_sfl_sld] offset:29728 \n"
" ds_read_b32 v16, %[v_sfl_sld] offset:29760 \n"
" ds_read_b32 v17, %[v_sfl_sld] offset:29792 \n"
" s_waitcnt lgkmcnt(0) \n"
" s_mov_b64 exec, %[s_execflag_0] \n" _UK_ATOMIC_ADD_
" %[v_os_o0], v10, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_1] \n" _UK_ATOMIC_ADD_
" %[v_os_o1], v11, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_2] \n" _UK_ATOMIC_ADD_
" %[v_os_o2], v12, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_3] \n" _UK_ATOMIC_ADD_
" %[v_os_o3], v13, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_4] \n" _UK_ATOMIC_ADD_
" %[v_os_o4], v14, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_5] \n" _UK_ATOMIC_ADD_
" %[v_os_o5], v15, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_6] \n" _UK_ATOMIC_ADD_
" %[v_os_o6], v16, s[8:9] \n"
" s_mov_b64 exec, %[s_execflag_7] \n" _UK_ATOMIC_ADD_
" %[v_os_o7], v17, s[8:9] \n"
" s_mov_b64 exec, s[38:39] \n"
#undef _UK_MFMA_
#undef _UK_MFMA_
#undef _UK_PK_CVT_
#undef _UK_ATOMIC_ADD_
include/ck_tile/ops/flatmm/block/uk/flatmm_uk_gfx9_32x512x128_1x1x1_16x16x16.inc
View file @ 2a768425
......@@ -9,508 +9,509 @@
#endif
"s_mov_b32 s16, %[s_res_a0] \n"
"s_mov_b32 s17, %[s_res_a1] \n"
"s_mov_b32 s18, %[s_res_a2] \n"
"s_mov_b32 s19, %[s_res_a3] \n"
"s_mov_b32 s20, %[s_res_b0] \n"
"s_mov_b32 s21, %[s_res_b1] \n"
"s_mov_b32 s22, %[s_res_b2] \n"
"s_mov_b32 s23, %[s_res_b3] \n"
// "s_nop 4\n"
"; -- prefetch A0\n"
"s_add_u32 m0, 0, %[s_m0_init] \n"
"buffer_load_dword %[v_os_a0], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a1], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a2], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a3], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a4], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a5], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a6], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a7], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[smem_sz], %[s_m0_init] \n"
"s_cmp_gt_i32 %[s_loop_cnt] 1 ; move a with cond \n"
"s_cselect_b32 s86, %[s_tile_os_a], 0 ; move a with cond \n"
"s_add_u32 s16, s86, s16 ; move a with cond \n"
"s_addc_u32 s17, 0, s17 ; move a with cond \n"
"; -- prefetch A1\n"
"buffer_load_dword %[v_os_a0], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a1], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a2], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a3], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a4], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a5], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a6], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a7], s[16:19], 0 offen lds \n"
"s_add_u32 m0, 0, %[s_m0_init] \n"
"s_cmp_gt_i32 %[s_loop_cnt] 2 ; move a with cond \n"
"s_cselect_b32 s86, %[s_tile_os_a], 0 ; move a with cond \n"
"s_add_u32 s16, s86, s16 ; move a with cond \n"
"s_addc_u32 s17, 0, s17 ; move a with cond \n"
"; -- prefetch B0\n"
"buffer_load_dwordx4 acc[0:3], %[v_os_b0], s[20:23], 0 offen \n"
"buffer_load_dwordx4 acc[4:7], %[v_os_b0], s[20:23], 0 offen offset:1024 \n"
"buffer_load_dwordx4 acc[8:11], %[v_os_b0], s[20:23], 0 offen offset:2048 \n"
"buffer_load_dwordx4 acc[12:15], %[v_os_b0], s[20:23], 0 offen offset:3072 \n"
"buffer_load_dwordx4 acc[16:19], %[v_os_b1], s[20:23], 0 offen \n"
"buffer_load_dwordx4 acc[20:23], %[v_os_b1], s[20:23], 0 offen offset:1024 \n"
"buffer_load_dwordx4 acc[24:27], %[v_os_b1], s[20:23], 0 offen offset:2048 \n"
"buffer_load_dwordx4 acc[28:31], %[v_os_b1], s[20:23], 0 offen offset:3072 \n"
"buffer_load_dwordx4 acc[32:35], %[v_os_b2], s[20:23], 0 offen \n"
"buffer_load_dwordx4 acc[36:39], %[v_os_b2], s[20:23], 0 offen offset:1024 \n"
"buffer_load_dwordx4 acc[40:43], %[v_os_b2], s[20:23], 0 offen offset:2048 \n"
"buffer_load_dwordx4 acc[44:47], %[v_os_b2], s[20:23], 0 offen offset:3072 \n"
"buffer_load_dwordx4 acc[48:51], %[v_os_b3], s[20:23], 0 offen \n"
"buffer_load_dwordx4 acc[52:55], %[v_os_b3], s[20:23], 0 offen offset:1024 \n"
"buffer_load_dwordx4 acc[56:59], %[v_os_b3], s[20:23], 0 offen offset:2048 \n"
"buffer_load_dwordx4 acc[60:63], %[v_os_b3], s[20:23], 0 offen offset:3072 \n"
"buffer_load_dwordx4 acc[64:67], %[v_os_b4], s[20:23], 0 offen \n"
"buffer_load_dwordx4 acc[68:71], %[v_os_b4], s[20:23], 0 offen offset:1024 \n"
"buffer_load_dwordx4 acc[72:75], %[v_os_b4], s[20:23], 0 offen offset:2048 \n"
"buffer_load_dwordx4 acc[76:79], %[v_os_b4], s[20:23], 0 offen offset:3072 \n"
"buffer_load_dwordx4 acc[80:83], %[v_os_b5], s[20:23], 0 offen \n"
"buffer_load_dwordx4 acc[84:87], %[v_os_b5], s[20:23], 0 offen offset:1024 \n"
"buffer_load_dwordx4 acc[88:91], %[v_os_b5], s[20:23], 0 offen offset:2048 \n"
"buffer_load_dwordx4 acc[92:95], %[v_os_b5], s[20:23], 0 offen offset:3072 \n"
"buffer_load_dwordx4 acc[96:99], %[v_os_b6], s[20:23], 0 offen \n"
"buffer_load_dwordx4 acc[100:103], %[v_os_b6], s[20:23], 0 offen offset:1024 \n"
"buffer_load_dwordx4 acc[104:107], %[v_os_b6], s[20:23], 0 offen offset:2048 \n"
"buffer_load_dwordx4 acc[108:111], %[v_os_b6], s[20:23], 0 offen offset:3072 \n"
"buffer_load_dwordx4 acc[112:115], %[v_os_b7], s[20:23], 0 offen \n"
"buffer_load_dwordx4 acc[116:119], %[v_os_b7], s[20:23], 0 offen offset:1024 \n"
"buffer_load_dwordx4 acc[120:123], %[v_os_b7], s[20:23], 0 offen offset:2048 \n"
"buffer_load_dwordx4 acc[124:127], %[v_os_b7], s[20:23], 0 offen offset:3072 \n"
"s_cmp_gt_i32 %[s_loop_cnt] 1 ; move b with cond \n"
"s_cselect_b32 s86, %[s_tile_os_b], 0 ; move b with cond \n"
"s_add_u32 s20, s86, s20 ; move b with cond \n"
"s_addc_u32 s21, 0, s21 ; move b with cond \n"
"s_waitcnt vmcnt(40) \n"
"s_barrier \n"
"ds_read_b128 v[64:67], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_0]\n" // 1024: N stride, 64 K stride
"ds_read_b128 v[68:71], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_1]\n"
"ds_read_b128 v[72:75], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_2]\n"
"ds_read_b128 v[76:79], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_3]\n"
"ds_read_b128 v[80:83], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_4]\n"
"ds_read_b128 v[84:87], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_5]\n"
"ds_read_b128 v[88:91], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_6]\n"
"ds_read_b128 v[92:95], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_7]\n"
"L_start%=: \n"
" s_waitcnt vmcnt(24) & lgkmcnt(0) \n"
" s_barrier \n"
_UK_MFMA_ " %[v_acc_0], acc[0:1], v[64:65], %[v_acc_0] \n"
_UK_MFMA_ " %[v_acc_0], acc[2:3], v[66:67], %[v_acc_0] \n"
" buffer_load_dwordx4 acc[128:131], %[v_os_b0], s[20:23], 0 offen \n"
_UK_MFMA_ " %[v_acc_0], acc[4:5], v[68:69], %[v_acc_0] \n"
_UK_MFMA_ " %[v_acc_0], acc[6:7], v[70:71], %[v_acc_0] \n"
" buffer_load_dword %[v_os_a0], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n"
_UK_MFMA_ " %[v_acc_0], acc[8:9], v[72:73], %[v_acc_0] \n"
_UK_MFMA_ " %[v_acc_0], acc[10:11], v[74:75], %[v_acc_0] \n"
" buffer_load_dwordx4 acc[132:135], %[v_os_b0], s[20:23], 0 offen offset:1024 \n"
_UK_MFMA_ " %[v_acc_0], acc[12:13], v[76:77], %[v_acc_0] \n"
_UK_MFMA_ " %[v_acc_0], acc[14:15], v[78:79], %[v_acc_0] \n"
" buffer_load_dword %[v_os_a1], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n"
_UK_MFMA_ " %[v_acc_1], acc[0:1], v[80:81], %[v_acc_1] \n"
_UK_MFMA_ " %[v_acc_1], acc[2:3], v[82:83], %[v_acc_1] \n"
" buffer_load_dwordx4 acc[136:139], %[v_os_b0], s[20:23], 0 offen offset:2048 \n"
_UK_MFMA_ " %[v_acc_1], acc[4:5], v[84:85], %[v_acc_1] \n"
_UK_MFMA_ " %[v_acc_1], acc[6:7], v[86:87], %[v_acc_1] \n"
" buffer_load_dword %[v_os_a2], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n"
_UK_MFMA_ " %[v_acc_1], acc[8:9], v[88:89], %[v_acc_1] \n"
_UK_MFMA_ " %[v_acc_1], acc[10:11], v[90:91], %[v_acc_1] \n"
" buffer_load_dwordx4 acc[140:143], %[v_os_b0], s[20:23], 0 offen offset:3072 \n"
_UK_MFMA_ " %[v_acc_1], acc[12:13], v[92:93], %[v_acc_1] \n"
_UK_MFMA_ " %[v_acc_1], acc[14:15], v[94:95], %[v_acc_1] \n"
" buffer_load_dword %[v_os_a3], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n"
_UK_MFMA_ " %[v_acc_2], acc[16:17], v[64:65], %[v_acc_2] \n"
_UK_MFMA_ " %[v_acc_2], acc[18:19], v[66:67], %[v_acc_2] \n"
" buffer_load_dwordx4 acc[144:147], %[v_os_b1], s[20:23], 0 offen \n"
_UK_MFMA_ " %[v_acc_2], acc[20:21], v[68:69], %[v_acc_2] \n"
_UK_MFMA_ " %[v_acc_2], acc[22:23], v[70:71], %[v_acc_2] \n"
" buffer_load_dword %[v_os_a4], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n"
_UK_MFMA_ " %[v_acc_2], acc[24:25], v[72:73], %[v_acc_2] \n"
_UK_MFMA_ " %[v_acc_2], acc[26:27], v[74:75], %[v_acc_2] \n"
" buffer_load_dwordx4 acc[148:151], %[v_os_b1], s[20:23], 0 offen offset:1024 \n"
_UK_MFMA_ " %[v_acc_2], acc[28:29], v[76:77], %[v_acc_2] \n"
_UK_MFMA_ " %[v_acc_2], acc[30:31], v[78:79], %[v_acc_2] \n"
" buffer_load_dword %[v_os_a5], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n"
_UK_MFMA_ " %[v_acc_3], acc[16:17], v[80:81], %[v_acc_3] \n"
_UK_MFMA_ " %[v_acc_3], acc[18:19], v[82:83], %[v_acc_3] \n"
" buffer_load_dwordx4 acc[152:155], %[v_os_b1], s[20:23], 0 offen offset:2048 \n"
_UK_MFMA_ " %[v_acc_3], acc[20:21], v[84:85], %[v_acc_3] \n"
_UK_MFMA_ " %[v_acc_3], acc[22:23], v[86:87], %[v_acc_3] \n"
" buffer_load_dword %[v_os_a6], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n"
_UK_MFMA_ " %[v_acc_3], acc[24:25], v[88:89], %[v_acc_3] \n"
_UK_MFMA_ " %[v_acc_3], acc[26:27], v[90:91], %[v_acc_3] \n"
" buffer_load_dwordx4 acc[156:159], %[v_os_b1], s[20:23], 0 offen offset:3072 \n"
_UK_MFMA_ " %[v_acc_3], acc[28:29], v[92:93], %[v_acc_3] \n"
_UK_MFMA_ " %[v_acc_3], acc[30:31], v[94:95], %[v_acc_3] \n"
" buffer_load_dword %[v_os_a7], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[smem_sz], %[s_m0_init] \n"
" s_waitcnt vmcnt(32) \n"
_UK_MFMA_ " %[v_acc_4], acc[32:33], v[64:65], %[v_acc_4] \n"
_UK_MFMA_ " %[v_acc_4], acc[34:35], v[66:67], %[v_acc_4] \n"
" buffer_load_dwordx4 acc[160:163], %[v_os_b2], s[20:23], 0 offen \n"
_UK_MFMA_ " %[v_acc_4], acc[36:37], v[68:69], %[v_acc_4] \n"
_UK_MFMA_ " %[v_acc_4], acc[38:39], v[70:71], %[v_acc_4] \n"
" ds_read_b128 v[96:99], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_0] \n"
_UK_MFMA_ " %[v_acc_4], acc[40:41], v[72:73], %[v_acc_4] \n"
_UK_MFMA_ " %[v_acc_4], acc[42:43], v[74:75], %[v_acc_4] \n"
" buffer_load_dwordx4 acc[164:167], %[v_os_b2], s[20:23], 0 offen offset:1024 \n"
_UK_MFMA_ " %[v_acc_4], acc[44:45], v[76:77], %[v_acc_4] \n"
_UK_MFMA_ " %[v_acc_4], acc[46:47], v[78:79], %[v_acc_4] \n"
" ds_read_b128 v[100:103], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_1] \n"
_UK_MFMA_ " %[v_acc_5], acc[32:33], v[80:81], %[v_acc_5] \n"
_UK_MFMA_ " %[v_acc_5], acc[34:35], v[82:83], %[v_acc_5] \n"
" buffer_load_dwordx4 acc[168:171], %[v_os_b2], s[20:23], 0 offen offset:2048 \n"
_UK_MFMA_ " %[v_acc_5], acc[36:37], v[84:85], %[v_acc_5] \n"
_UK_MFMA_ " %[v_acc_5], acc[38:39], v[86:87], %[v_acc_5] \n"
" ds_read_b128 v[104:107], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_2] \n"
_UK_MFMA_ " %[v_acc_5], acc[40:41], v[88:89], %[v_acc_5] \n"
_UK_MFMA_ " %[v_acc_5], acc[42:43], v[90:91], %[v_acc_5] \n"
" buffer_load_dwordx4 acc[172:175], %[v_os_b2], s[20:23], 0 offen offset:3072 \n"
_UK_MFMA_ " %[v_acc_5], acc[44:45], v[92:93], %[v_acc_5] \n"
_UK_MFMA_ " %[v_acc_5], acc[46:47], v[94:95], %[v_acc_5] \n"
" ds_read_b128 v[108:111], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_3] \n"
_UK_MFMA_ " %[v_acc_6], acc[48:49], v[64:65], %[v_acc_6] \n"
_UK_MFMA_ " %[v_acc_6], acc[50:51], v[66:67], %[v_acc_6] \n"
" buffer_load_dwordx4 acc[176:179], %[v_os_b3], s[20:23], 0 offen \n"
_UK_MFMA_ " %[v_acc_6], acc[52:53], v[68:69], %[v_acc_6] \n"
_UK_MFMA_ " %[v_acc_6], acc[54:55], v[70:71], %[v_acc_6] \n"
" ds_read_b128 v[112:115], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_4] \n"
_UK_MFMA_ " %[v_acc_6], acc[56:57], v[72:73], %[v_acc_6] \n"
_UK_MFMA_ " %[v_acc_6], acc[58:59], v[74:75], %[v_acc_6] \n"
" buffer_load_dwordx4 acc[180:183], %[v_os_b3], s[20:23], 0 offen offset:1024 \n"
_UK_MFMA_ " %[v_acc_6], acc[60:61], v[76:77], %[v_acc_6] \n"
_UK_MFMA_ " %[v_acc_6], acc[62:63], v[78:79], %[v_acc_6] \n"
" ds_read_b128 v[116:119], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_5] \n"
_UK_MFMA_ " %[v_acc_7], acc[48:49], v[80:81], %[v_acc_7] \n"
_UK_MFMA_ " %[v_acc_7], acc[50:51], v[82:83], %[v_acc_7] \n"
" buffer_load_dwordx4 acc[184:187], %[v_os_b3], s[20:23], 0 offen offset:2048 \n"
_UK_MFMA_ " %[v_acc_7], acc[52:53], v[84:85], %[v_acc_7] \n"
_UK_MFMA_ " %[v_acc_7], acc[54:55], v[86:87], %[v_acc_7] \n"
" ds_read_b128 v[120:123], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_6] \n"
_UK_MFMA_ " %[v_acc_7], acc[56:57], v[88:89], %[v_acc_7] \n"
_UK_MFMA_ " %[v_acc_7], acc[58:59], v[90:91], %[v_acc_7] \n"
" buffer_load_dwordx4 acc[188:191], %[v_os_b3], s[20:23], 0 offen offset:3072 \n"
_UK_MFMA_ " %[v_acc_7], acc[60:61], v[92:93], %[v_acc_7] \n"
_UK_MFMA_ " %[v_acc_7], acc[62:63], v[94:95], %[v_acc_7] \n"
" ds_read_b128 v[124:127], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_7] \n"
" s_waitcnt vmcnt(32) \n"
_UK_MFMA_ " %[v_acc_8], acc[64:65], v[64:65], %[v_acc_8] \n"
_UK_MFMA_ " %[v_acc_8], acc[66:67], v[66:67], %[v_acc_8] \n"
" buffer_load_dwordx4 acc[192:195], %[v_os_b4], s[20:23], 0 offen \n"
_UK_MFMA_ " %[v_acc_8], acc[68:69], v[68:69], %[v_acc_8] \n"
_UK_MFMA_ " %[v_acc_8], acc[70:71], v[70:71], %[v_acc_8] \n"
_UK_MFMA_ " %[v_acc_8], acc[72:73], v[72:73], %[v_acc_8] \n"
_UK_MFMA_ " %[v_acc_8], acc[74:75], v[74:75], %[v_acc_8] \n"
" buffer_load_dwordx4 acc[196:199], %[v_os_b4], s[20:23], 0 offen offset:1024 \n"
_UK_MFMA_ " %[v_acc_8], acc[76:77], v[76:77], %[v_acc_8] \n"
_UK_MFMA_ " %[v_acc_8], acc[78:79], v[78:79], %[v_acc_8] \n"
_UK_MFMA_ " %[v_acc_9], acc[64:65], v[80:81], %[v_acc_9] \n"
_UK_MFMA_ " %[v_acc_9], acc[66:67], v[82:83], %[v_acc_9] \n"
" buffer_load_dwordx4 acc[200:203], %[v_os_b4], s[20:23], 0 offen offset:2048 \n"
_UK_MFMA_ " %[v_acc_9], acc[68:69], v[84:85], %[v_acc_9] \n"
_UK_MFMA_ " %[v_acc_9], acc[70:71], v[86:87], %[v_acc_9] \n"
_UK_MFMA_ " %[v_acc_9], acc[72:73], v[88:89], %[v_acc_9] \n"
_UK_MFMA_ " %[v_acc_9], acc[74:75], v[90:91], %[v_acc_9] \n"
" buffer_load_dwordx4 acc[204:207], %[v_os_b4], s[20:23], 0 offen offset:3072 \n"
_UK_MFMA_ " %[v_acc_9], acc[76:77], v[92:93], %[v_acc_9] \n"
_UK_MFMA_ " %[v_acc_9], acc[78:79], v[94:95], %[v_acc_9] \n"
_UK_MFMA_ " %[v_acc_10], acc[80:81], v[64:65], %[v_acc_10] \n"
_UK_MFMA_ " %[v_acc_10], acc[82:83], v[66:67], %[v_acc_10] \n"
" buffer_load_dwordx4 acc[208:211], %[v_os_b5], s[20:23], 0 offen \n"
_UK_MFMA_ " %[v_acc_10], acc[84:85], v[68:69], %[v_acc_10] \n"
_UK_MFMA_ " %[v_acc_10], acc[86:87], v[70:71], %[v_acc_10] \n"
_UK_MFMA_ " %[v_acc_10], acc[88:89], v[72:73], %[v_acc_10] \n"
_UK_MFMA_ " %[v_acc_10], acc[90:91], v[74:75], %[v_acc_10] \n"
" buffer_load_dwordx4 acc[212:215], %[v_os_b5], s[20:23], 0 offen offset:1024 \n"
_UK_MFMA_ " %[v_acc_10], acc[92:93], v[76:77], %[v_acc_10] \n"
_UK_MFMA_ " %[v_acc_10], acc[94:95], v[78:79], %[v_acc_10] \n"
_UK_MFMA_ " %[v_acc_11], acc[80:81], v[80:81], %[v_acc_11] \n"
_UK_MFMA_ " %[v_acc_11], acc[82:83], v[82:83], %[v_acc_11] \n"
" buffer_load_dwordx4 acc[216:219], %[v_os_b5], s[20:23], 0 offen offset:2048 \n"
_UK_MFMA_ " %[v_acc_11], acc[84:85], v[84:85], %[v_acc_11] \n"
_UK_MFMA_ " %[v_acc_11], acc[86:87], v[86:87], %[v_acc_11] \n"
_UK_MFMA_ " %[v_acc_11], acc[88:89], v[88:89], %[v_acc_11] \n"
_UK_MFMA_ " %[v_acc_11], acc[90:91], v[90:91], %[v_acc_11] \n"
" buffer_load_dwordx4 acc[220:223], %[v_os_b5], s[20:23], 0 offen offset:3072 \n"
_UK_MFMA_ " %[v_acc_11], acc[92:93], v[92:93], %[v_acc_11] \n"
_UK_MFMA_ " %[v_acc_11], acc[94:95], v[94:95], %[v_acc_11] \n"
" s_waitcnt vmcnt(32) \n"
_UK_MFMA_ " %[v_acc_12], acc[96:97], v[64:65], %[v_acc_12] \n"
_UK_MFMA_ " %[v_acc_12], acc[98:99], v[66:67], %[v_acc_12] \n"
" buffer_load_dwordx4 acc[224:227], %[v_os_b6], s[20:23], 0 offen \n"
_UK_MFMA_ " %[v_acc_12], acc[100:101], v[68:69], %[v_acc_12] \n"
_UK_MFMA_ " %[v_acc_12], acc[102:103], v[70:71], %[v_acc_12] \n"
_UK_MFMA_ " %[v_acc_12], acc[104:105], v[72:73], %[v_acc_12] \n"
_UK_MFMA_ " %[v_acc_12], acc[106:107], v[74:75], %[v_acc_12] \n"
" buffer_load_dwordx4 acc[228:231], %[v_os_b6], s[20:23], 0 offen offset:1024 \n"
_UK_MFMA_ " %[v_acc_12], acc[108:109], v[76:77], %[v_acc_12] \n"
_UK_MFMA_ " %[v_acc_12], acc[110:111], v[78:79], %[v_acc_12] \n"
_UK_MFMA_ " %[v_acc_13], acc[96:97], v[80:81], %[v_acc_13] \n"
_UK_MFMA_ " %[v_acc_13], acc[98:99], v[82:83], %[v_acc_13] \n"
" buffer_load_dwordx4 acc[232:235], %[v_os_b6], s[20:23], 0 offen offset:2048 \n"
_UK_MFMA_ " %[v_acc_13], acc[100:101], v[84:85], %[v_acc_13] \n"
_UK_MFMA_ " %[v_acc_13], acc[102:103], v[86:87], %[v_acc_13] \n"
_UK_MFMA_ " %[v_acc_13], acc[104:105], v[88:89], %[v_acc_13] \n"
_UK_MFMA_ " %[v_acc_13], acc[106:107], v[90:91], %[v_acc_13] \n"
" buffer_load_dwordx4 acc[236:239], %[v_os_b6], s[20:23], 0 offen offset:3072 \n"
_UK_MFMA_ " %[v_acc_13], acc[108:109], v[92:93], %[v_acc_13] \n"
_UK_MFMA_ " %[v_acc_13], acc[110:111], v[94:95], %[v_acc_13] \n"
_UK_MFMA_ " %[v_acc_14], acc[112:113], v[64:65], %[v_acc_14] \n"
_UK_MFMA_ " %[v_acc_14], acc[114:115], v[66:67], %[v_acc_14] \n"
" buffer_load_dwordx4 acc[240:243], %[v_os_b7], s[20:23], 0 offen \n"
_UK_MFMA_ " %[v_acc_14], acc[116:117], v[68:69], %[v_acc_14] \n"
_UK_MFMA_ " %[v_acc_14], acc[118:119], v[70:71], %[v_acc_14] \n"
_UK_MFMA_ " %[v_acc_14], acc[120:121], v[72:73], %[v_acc_14] \n"
_UK_MFMA_ " %[v_acc_14], acc[122:123], v[74:75], %[v_acc_14] \n"
" buffer_load_dwordx4 acc[244:247], %[v_os_b7], s[20:23], 0 offen offset:1024 \n"
_UK_MFMA_ " %[v_acc_14], acc[124:125], v[76:77], %[v_acc_14] \n"
_UK_MFMA_ " %[v_acc_14], acc[126:127], v[78:79], %[v_acc_14] \n"
_UK_MFMA_ " %[v_acc_15], acc[112:113], v[80:81], %[v_acc_15] \n"
_UK_MFMA_ " %[v_acc_15], acc[114:115], v[82:83], %[v_acc_15] \n"
" buffer_load_dwordx4 acc[248:251], %[v_os_b7], s[20:23], 0 offen offset:2048 \n"
_UK_MFMA_ " %[v_acc_15], acc[116:117], v[84:85], %[v_acc_15] \n"
_UK_MFMA_ " %[v_acc_15], acc[118:119], v[86:87], %[v_acc_15] \n"
_UK_MFMA_ " %[v_acc_15], acc[120:121], v[88:89], %[v_acc_15] \n"
_UK_MFMA_ " %[v_acc_15], acc[122:123], v[90:91], %[v_acc_15] \n"
" buffer_load_dwordx4 acc[252:255], %[v_os_b7], s[20:23], 0 offen offset:3072\n"
_UK_MFMA_ " %[v_acc_15], acc[124:125], v[92:93], %[v_acc_15] \n"
_UK_MFMA_ " %[v_acc_15], acc[126:127], v[94:95], %[v_acc_15] \n"
" s_sub_i32 %[s_loop_cnt], %[s_loop_cnt], 1 \n"
" s_cmp_gt_i32 %[s_loop_cnt] 0 \n"
" s_cbranch_scc0 L_end%= \n"
" s_cmp_gt_i32 %[s_loop_cnt] 2 ; move a with cond \n"
" s_cselect_b32 s86, %[s_tile_os_a], 0 \n"
" s_add_u32 s16, s86, s16 \n"
" s_addc_u32 s17, 0, s17 \n"
" s_cmp_gt_i32 %[s_loop_cnt] 1 ; move b with cond \n"
" s_cselect_b32 s86, %[s_tile_os_b], 0 \n"
" s_add_u32 s20, s86, s20 \n"
" s_addc_u32 s21, 0, s21 \n"
" ;------------------------------------------ \n"
" s_waitcnt vmcnt(24) & lgkmcnt(0) \n"
" s_barrier \n"
_UK_MFMA_ " %[v_acc_0], acc[128:129], v[96:97], %[v_acc_0] \n"
_UK_MFMA_ " %[v_acc_0], acc[130:131], v[98:99], %[v_acc_0] \n"
" buffer_load_dwordx4 acc[0:3], %[v_os_b0], s[20:23], 0 offen \n"
_UK_MFMA_ " %[v_acc_0], acc[132:133], v[100:101], %[v_acc_0] \n"
_UK_MFMA_ " %[v_acc_0], acc[134:135], v[102:103], %[v_acc_0] \n"
" buffer_load_dword %[v_os_a0], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n"
_UK_MFMA_ " %[v_acc_0], acc[136:137], v[104:105], %[v_acc_0] \n"
_UK_MFMA_ " %[v_acc_0], acc[138:139], v[106:107], %[v_acc_0] \n"
" buffer_load_dwordx4 acc[4:7], %[v_os_b0], s[20:23], 0 offen offset:1024 \n"
_UK_MFMA_ " %[v_acc_0], acc[140:141], v[108:109], %[v_acc_0] \n"
_UK_MFMA_ " %[v_acc_0], acc[142:143], v[110:111], %[v_acc_0] \n"
" buffer_load_dword %[v_os_a1], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n"
_UK_MFMA_ " %[v_acc_1], acc[128:129], v[112:113], %[v_acc_1] \n"
_UK_MFMA_ " %[v_acc_1], acc[130:131], v[114:115], %[v_acc_1] \n"
" buffer_load_dwordx4 acc[8:11], %[v_os_b0], s[20:23], 0 offen offset:2048 \n"
_UK_MFMA_ " %[v_acc_1], acc[132:133], v[116:117], %[v_acc_1] \n"
_UK_MFMA_ " %[v_acc_1], acc[134:135], v[118:119], %[v_acc_1] \n"
" buffer_load_dword %[v_os_a2], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n"
_UK_MFMA_ " %[v_acc_1], acc[136:137], v[120:121], %[v_acc_1] \n"
_UK_MFMA_ " %[v_acc_1], acc[138:139], v[122:123], %[v_acc_1] \n"
" buffer_load_dwordx4 acc[12:15], %[v_os_b0], s[20:23], 0 offen offset:3072 \n"
_UK_MFMA_ " %[v_acc_1], acc[140:141], v[124:125], %[v_acc_1] \n"
_UK_MFMA_ " %[v_acc_1], acc[142:143], v[126:127], %[v_acc_1] \n"
" buffer_load_dword %[v_os_a3], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n"
_UK_MFMA_ " %[v_acc_2], acc[144:145], v[96:97], %[v_acc_2] \n"
_UK_MFMA_ " %[v_acc_2], acc[146:147], v[98:99], %[v_acc_2] \n"
" buffer_load_dwordx4 acc[16:19], %[v_os_b1], s[20:23], 0 offen \n"
_UK_MFMA_ " %[v_acc_2], acc[148:149], v[100:101], %[v_acc_2] \n"
_UK_MFMA_ " %[v_acc_2], acc[150:151], v[102:103], %[v_acc_2] \n"
" buffer_load_dword %[v_os_a4], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n"
_UK_MFMA_ " %[v_acc_2], acc[152:153], v[104:105], %[v_acc_2] \n"
_UK_MFMA_ " %[v_acc_2], acc[154:155], v[106:107], %[v_acc_2] \n"
" buffer_load_dwordx4 acc[20:23], %[v_os_b1], s[20:23], 0 offen offset:1024 \n"
_UK_MFMA_ " %[v_acc_2], acc[156:157], v[108:109], %[v_acc_2] \n"
_UK_MFMA_ " %[v_acc_2], acc[158:159], v[110:111], %[v_acc_2] \n"
" buffer_load_dword %[v_os_a5], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n"
_UK_MFMA_ " %[v_acc_3], acc[144:145], v[112:113], %[v_acc_3] \n"
_UK_MFMA_ " %[v_acc_3], acc[146:147], v[114:115], %[v_acc_3] \n"
" buffer_load_dwordx4 acc[24:27], %[v_os_b1], s[20:23], 0 offen offset:2048 \n"
_UK_MFMA_ " %[v_acc_3], acc[148:149], v[116:117], %[v_acc_3] \n"
_UK_MFMA_ " %[v_acc_3], acc[150:151], v[118:119], %[v_acc_3] \n"
" buffer_load_dword %[v_os_a6], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n"
_UK_MFMA_ " %[v_acc_3], acc[152:153], v[120:121], %[v_acc_3] \n"
_UK_MFMA_ " %[v_acc_3], acc[154:155], v[122:123], %[v_acc_3] \n"
" buffer_load_dwordx4 acc[28:31], %[v_os_b1], s[20:23], 0 offen offset:3072 \n"
_UK_MFMA_ " %[v_acc_3], acc[156:157], v[124:125], %[v_acc_3] \n"
_UK_MFMA_ " %[v_acc_3], acc[158:159], v[126:127], %[v_acc_3] \n"
" buffer_load_dword %[v_os_a7], s[16:19], 0 offen lds \n"
" s_add_u32 m0, 0, %[s_m0_init] \n"
" s_waitcnt vmcnt(32) \n"
_UK_MFMA_ " %[v_acc_4], acc[160:161], v[96:97], %[v_acc_4] \n"
_UK_MFMA_ " %[v_acc_4], acc[162:163], v[98:99], %[v_acc_4] \n"
" buffer_load_dwordx4 acc[32:35], %[v_os_b2], s[20:23], 0 offen \n"
_UK_MFMA_ " %[v_acc_4], acc[164:165], v[100:101], %[v_acc_4] \n"
_UK_MFMA_ " %[v_acc_4], acc[166:167], v[102:103], %[v_acc_4] \n"
" ds_read_b128 v[64:67], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_0] \n"
_UK_MFMA_ " %[v_acc_4], acc[168:169], v[104:105], %[v_acc_4] \n"
_UK_MFMA_ " %[v_acc_4], acc[170:171], v[106:107], %[v_acc_4] \n"
" buffer_load_dwordx4 acc[36:39], %[v_os_b2], s[20:23], 0 offen offset:1024 \n"
_UK_MFMA_ " %[v_acc_4], acc[172:173], v[108:109], %[v_acc_4] \n"
_UK_MFMA_ " %[v_acc_4], acc[174:175], v[110:111], %[v_acc_4] \n"
" ds_read_b128 v[68:71], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_1] \n"
_UK_MFMA_ " %[v_acc_5], acc[160:161], v[112:113], %[v_acc_5] \n"
_UK_MFMA_ " %[v_acc_5], acc[162:163], v[114:115], %[v_acc_5] \n"
" buffer_load_dwordx4 acc[40:43], %[v_os_b2], s[20:23], 0 offen offset:2048 \n"
_UK_MFMA_ " %[v_acc_5], acc[164:165], v[116:117], %[v_acc_5] \n"
_UK_MFMA_ " %[v_acc_5], acc[166:167], v[118:119], %[v_acc_5] \n"
" ds_read_b128 v[72:75], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_2] \n"
_UK_MFMA_ " %[v_acc_5], acc[168:169], v[120:121], %[v_acc_5] \n"
_UK_MFMA_ " %[v_acc_5], acc[170:171], v[122:123], %[v_acc_5] \n"
" buffer_load_dwordx4 acc[44:47], %[v_os_b2], s[20:23], 0 offen offset:3072 \n"
_UK_MFMA_ " %[v_acc_5], acc[172:173], v[124:125], %[v_acc_5] \n"
_UK_MFMA_ " %[v_acc_5], acc[174:175], v[126:127], %[v_acc_5] \n"
" ds_read_b128 v[76:79], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_3] \n"
_UK_MFMA_ " %[v_acc_6], acc[176:177], v[96:97], %[v_acc_6] \n"
_UK_MFMA_ " %[v_acc_6], acc[178:179], v[98:99], %[v_acc_6] \n"
" buffer_load_dwordx4 acc[48:51], %[v_os_b3], s[20:23], 0 offen \n"
_UK_MFMA_ " %[v_acc_6], acc[180:181], v[100:101], %[v_acc_6] \n"
_UK_MFMA_ " %[v_acc_6], acc[182:183], v[102:103], %[v_acc_6] \n"
" ds_read_b128 v[80:83], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_4] \n"
_UK_MFMA_ " %[v_acc_6], acc[184:185], v[104:105], %[v_acc_6] \n"
_UK_MFMA_ " %[v_acc_6], acc[186:187], v[106:107], %[v_acc_6] \n"
" buffer_load_dwordx4 acc[52:55], %[v_os_b3], s[20:23], 0 offen offset:1024 \n"
_UK_MFMA_ " %[v_acc_6], acc[188:189], v[108:109], %[v_acc_6] \n"
_UK_MFMA_ " %[v_acc_6], acc[190:191], v[110:111], %[v_acc_6] \n"
" ds_read_b128 v[84:87], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_5] \n"
_UK_MFMA_ " %[v_acc_7], acc[176:177], v[112:113], %[v_acc_7] \n"
_UK_MFMA_ " %[v_acc_7], acc[178:179], v[114:115], %[v_acc_7] \n"
" buffer_load_dwordx4 acc[56:59], %[v_os_b3], s[20:23], 0 offen offset:2048 \n"
_UK_MFMA_ " %[v_acc_7], acc[180:181], v[116:117], %[v_acc_7] \n"
_UK_MFMA_ " %[v_acc_7], acc[182:183], v[118:119], %[v_acc_7] \n"
" ds_read_b128 v[88:91], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_6] \n"
_UK_MFMA_ " %[v_acc_7], acc[184:185], v[120:121], %[v_acc_7] \n"
_UK_MFMA_ " %[v_acc_7], acc[186:187], v[122:123], %[v_acc_7] \n"
" buffer_load_dwordx4 acc[60:63], %[v_os_b3], s[20:23], 0 offen offset:3072 \n"
_UK_MFMA_ " %[v_acc_7], acc[188:189], v[124:125], %[v_acc_7] \n"
_UK_MFMA_ " %[v_acc_7], acc[190:191], v[126:127], %[v_acc_7] \n"
" ds_read_b128 v[92:95], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_7] \n"
" s_waitcnt vmcnt(32) \n"
_UK_MFMA_ " %[v_acc_8], acc[192:193], v[96:97], %[v_acc_8] \n"
_UK_MFMA_ " %[v_acc_8], acc[194:195], v[98:99], %[v_acc_8] \n"
" buffer_load_dwordx4 acc[64:67], %[v_os_b4], s[20:23], 0 offen \n"
_UK_MFMA_ " %[v_acc_8], acc[196:197], v[100:101], %[v_acc_8] \n"
_UK_MFMA_ " %[v_acc_8], acc[198:199], v[102:103], %[v_acc_8] \n"
_UK_MFMA_ " %[v_acc_8], acc[200:201], v[104:105], %[v_acc_8] \n"
_UK_MFMA_ " %[v_acc_8], acc[202:203], v[106:107], %[v_acc_8] \n"
" buffer_load_dwordx4 acc[68:71], %[v_os_b4], s[20:23], 0 offen offset:1024 \n"
_UK_MFMA_ " %[v_acc_8], acc[204:205], v[108:109], %[v_acc_8] \n"
_UK_MFMA_ " %[v_acc_8], acc[206:207], v[110:111], %[v_acc_8] \n"
_UK_MFMA_ " %[v_acc_9], acc[192:193], v[112:113], %[v_acc_9] \n"
_UK_MFMA_ " %[v_acc_9], acc[194:195], v[114:115], %[v_acc_9] \n"
" buffer_load_dwordx4 acc[72:75], %[v_os_b4], s[20:23], 0 offen offset:2048 \n"
_UK_MFMA_ " %[v_acc_9], acc[196:197], v[116:117], %[v_acc_9] \n"
_UK_MFMA_ " %[v_acc_9], acc[198:199], v[118:119], %[v_acc_9] \n"
_UK_MFMA_ " %[v_acc_9], acc[200:201], v[120:121], %[v_acc_9] \n"
_UK_MFMA_ " %[v_acc_9], acc[202:203], v[122:123], %[v_acc_9] \n"
" buffer_load_dwordx4 acc[76:79], %[v_os_b4], s[20:23], 0 offen offset:3072 \n"
_UK_MFMA_ " %[v_acc_9], acc[204:205], v[124:125], %[v_acc_9] \n"
_UK_MFMA_ " %[v_acc_9], acc[206:207], v[126:127], %[v_acc_9] \n"
_UK_MFMA_ " %[v_acc_10], acc[208:209], v[96:97], %[v_acc_10] \n"
_UK_MFMA_ " %[v_acc_10], acc[210:211], v[98:99], %[v_acc_10] \n"
" buffer_load_dwordx4 acc[80:83], %[v_os_b5], s[20:23], 0 offen \n"
_UK_MFMA_ " %[v_acc_10], acc[212:213], v[100:101], %[v_acc_10] \n"
_UK_MFMA_ " %[v_acc_10], acc[214:215], v[102:103], %[v_acc_10] \n"
_UK_MFMA_ " %[v_acc_10], acc[216:217], v[104:105], %[v_acc_10] \n"
_UK_MFMA_ " %[v_acc_10], acc[218:219], v[106:107], %[v_acc_10] \n"
" buffer_load_dwordx4 acc[84:87], %[v_os_b5], s[20:23], 0 offen offset:1024 \n"
_UK_MFMA_ " %[v_acc_10], acc[220:221], v[108:109], %[v_acc_10] \n"
_UK_MFMA_ " %[v_acc_10], acc[222:223], v[110:111], %[v_acc_10] \n"
_UK_MFMA_ " %[v_acc_11], acc[208:209], v[112:113], %[v_acc_11] \n"
_UK_MFMA_ " %[v_acc_11], acc[210:211], v[114:115], %[v_acc_11] \n"
" buffer_load_dwordx4 acc[88:91], %[v_os_b5], s[20:23], 0 offen offset:2048 \n"
_UK_MFMA_ " %[v_acc_11], acc[212:213], v[116:117], %[v_acc_11] \n"
_UK_MFMA_ " %[v_acc_11], acc[214:215], v[118:119], %[v_acc_11] \n"
_UK_MFMA_ " %[v_acc_11], acc[216:217], v[120:121], %[v_acc_11] \n"
_UK_MFMA_ " %[v_acc_11], acc[218:219], v[122:123], %[v_acc_11] \n"
" buffer_load_dwordx4 acc[92:95], %[v_os_b5], s[20:23], 0 offen offset:3072 \n"
_UK_MFMA_ " %[v_acc_11], acc[220:221], v[124:125], %[v_acc_11] \n"
_UK_MFMA_ " %[v_acc_11], acc[222:223], v[126:127], %[v_acc_11] \n"
" s_waitcnt vmcnt(32) \n"
_UK_MFMA_ " %[v_acc_12], acc[224:225], v[96:97], %[v_acc_12] \n"
_UK_MFMA_ " %[v_acc_12], acc[226:227], v[98:99], %[v_acc_12] \n"
" buffer_load_dwordx4 acc[96:99], %[v_os_b6], s[20:23], 0 offen \n"
_UK_MFMA_ " %[v_acc_12], acc[228:229], v[100:101], %[v_acc_12] \n"
_UK_MFMA_ " %[v_acc_12], acc[230:231], v[102:103], %[v_acc_12] \n"
_UK_MFMA_ " %[v_acc_12], acc[232:233], v[104:105], %[v_acc_12] \n"
_UK_MFMA_ " %[v_acc_12], acc[234:235], v[106:107], %[v_acc_12] \n"
" buffer_load_dwordx4 acc[100:103], %[v_os_b6], s[20:23], 0 offen offset:1024 \n"
_UK_MFMA_ " %[v_acc_12], acc[236:237], v[108:109], %[v_acc_12] \n"
_UK_MFMA_ " %[v_acc_12], acc[238:239], v[110:111], %[v_acc_12] \n"
_UK_MFMA_ " %[v_acc_13], acc[224:225], v[112:113], %[v_acc_13] \n"
_UK_MFMA_ " %[v_acc_13], acc[226:227], v[114:115], %[v_acc_13] \n"
" buffer_load_dwordx4 acc[104:107], %[v_os_b6], s[20:23], 0 offen offset:2048 \n"
_UK_MFMA_ " %[v_acc_13], acc[228:229], v[116:117], %[v_acc_13] \n"
_UK_MFMA_ " %[v_acc_13], acc[230:231], v[118:119], %[v_acc_13] \n"
_UK_MFMA_ " %[v_acc_13], acc[232:233], v[120:121], %[v_acc_13] \n"
_UK_MFMA_ " %[v_acc_13], acc[234:235], v[122:123], %[v_acc_13] \n"
" buffer_load_dwordx4 acc[108:111], %[v_os_b6], s[20:23], 0 offen offset:3072 \n"
_UK_MFMA_ " %[v_acc_13], acc[236:237], v[124:125], %[v_acc_13] \n"
_UK_MFMA_ " %[v_acc_13], acc[238:239], v[126:127], %[v_acc_13] \n"
_UK_MFMA_ " %[v_acc_14], acc[240:241], v[96:97], %[v_acc_14] \n"
_UK_MFMA_ " %[v_acc_14], acc[242:243], v[98:99], %[v_acc_14] \n"
" buffer_load_dwordx4 acc[112:115], %[v_os_b7], s[20:23], 0 offen \n"
_UK_MFMA_ " %[v_acc_14], acc[244:245], v[100:101], %[v_acc_14] \n"
_UK_MFMA_ " %[v_acc_14], acc[246:247], v[102:103], %[v_acc_14] \n"
_UK_MFMA_ " %[v_acc_14], acc[248:249], v[104:105], %[v_acc_14] \n"
_UK_MFMA_ " %[v_acc_14], acc[250:251], v[106:107], %[v_acc_14] \n"
" buffer_load_dwordx4 acc[116:119], %[v_os_b7], s[20:23], 0 offen offset:1024 \n"
_UK_MFMA_ " %[v_acc_14], acc[252:253], v[108:109], %[v_acc_14] \n"
_UK_MFMA_ " %[v_acc_14], acc[254:255], v[110:111], %[v_acc_14] \n"
_UK_MFMA_ " %[v_acc_15], acc[240:241], v[112:113], %[v_acc_15] \n"
_UK_MFMA_ " %[v_acc_15], acc[242:243], v[114:115], %[v_acc_15] \n"
" buffer_load_dwordx4 acc[120:123], %[v_os_b7], s[20:23], 0 offen offset:2048 \n"
_UK_MFMA_ " %[v_acc_15], acc[244:245], v[116:117], %[v_acc_15] \n"
_UK_MFMA_ " %[v_acc_15], acc[246:247], v[118:119], %[v_acc_15] \n"
_UK_MFMA_ " %[v_acc_15], acc[248:249], v[120:121], %[v_acc_15] \n"
_UK_MFMA_ " %[v_acc_15], acc[250:251], v[122:123], %[v_acc_15] \n"
" buffer_load_dwordx4 acc[124:127], %[v_os_b7], s[20:23], 0 offen offset:3072 \n"
_UK_MFMA_ " %[v_acc_15], acc[252:253], v[124:125], %[v_acc_15] \n"
_UK_MFMA_ " %[v_acc_15], acc[254:255], v[126:127], %[v_acc_15] \n"
" s_sub_i32 %[s_loop_cnt], %[s_loop_cnt], 1 \n"
" s_cmp_gt_i32 %[s_loop_cnt] 0 \n"
" s_cbranch_scc0 L_end%= \n"
" s_cmp_gt_i32 %[s_loop_cnt] 2 ; move a with cond \n"
" s_cselect_b32 s86, %[s_tile_os_a], 0 \n"
" s_add_u32 s16, s86, s16 \n"
" s_addc_u32 s17, 0, s17 \n"
" s_cmp_gt_i32 %[s_loop_cnt] 1 ; move b with cond \n"
" s_cselect_b32 s86, %[s_tile_os_b], 0 \n"
" s_add_u32 s20, s86, s20 \n"
" s_addc_u32 s21, 0, s21 \n"
" s_branch L_start%= \n"
"L_end%=: \n"
" s_nop 2 \n"
"s_mov_b32 s17, %[s_res_a1] \n"
"s_mov_b32 s18, %[s_res_a2] \n"
"s_mov_b32 s19, %[s_res_a3] \n"
"s_mov_b32 s20, %[s_res_b0] \n"
"s_mov_b32 s21, %[s_res_b1] \n"
"s_mov_b32 s22, %[s_res_b2] \n"
"s_mov_b32 s23, %[s_res_b3] \n"
// "s_nop 4\n"
"; -- prefetch A0\n"
"s_add_u32 m0, 0, %[s_m0_init] \n"
"buffer_load_dword %[v_os_a0], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a1], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a2], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a3], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a4], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a5], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a6], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a7], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[smem_sz], %[s_m0_init] \n"
"s_cmp_gt_i32 %[s_loop_cnt] 1 ; move a with cond \n"
"s_cselect_b32 s86, %[s_tile_os_a], 0 ; move a with cond \n"
"s_add_u32 s16, s86, s16 ; move a with cond \n"
"s_addc_u32 s17, 0, s17 ; move a with cond \n"
"; -- prefetch A1\n"
"buffer_load_dword %[v_os_a0], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a1], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a2], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a3], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a4], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a5], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a6], s[16:19], 0 offen lds \n"
"s_add_u32 m0, %[s_size_per_issue], m0 \n"
"buffer_load_dword %[v_os_a7], s[16:19], 0 offen lds \n"
"s_add_u32 m0, 0, %[s_m0_init] \n"
"s_cmp_gt_i32 %[s_loop_cnt] 2 ; move a with cond \n"
"s_cselect_b32 s86, %[s_tile_os_a], 0 ; move a with cond \n"
"s_add_u32 s16, s86, s16 ; move a with cond \n"
"s_addc_u32 s17, 0, s17 ; move a with cond \n"
"; -- prefetch B0\n"
"buffer_load_dwordx4 acc[0:3], %[v_os_b0], s[20:23], 0 offen \n"
"buffer_load_dwordx4 acc[4:7], %[v_os_b0], s[20:23], 0 offen offset:1024 \n"
"buffer_load_dwordx4 acc[8:11], %[v_os_b0], s[20:23], 0 offen offset:2048 \n"
"buffer_load_dwordx4 acc[12:15], %[v_os_b0], s[20:23], 0 offen offset:3072 \n"
"buffer_load_dwordx4 acc[16:19], %[v_os_b1], s[20:23], 0 offen \n"
"buffer_load_dwordx4 acc[20:23], %[v_os_b1], s[20:23], 0 offen offset:1024 \n"
"buffer_load_dwordx4 acc[24:27], %[v_os_b1], s[20:23], 0 offen offset:2048 \n"
"buffer_load_dwordx4 acc[28:31], %[v_os_b1], s[20:23], 0 offen offset:3072 \n"
"buffer_load_dwordx4 acc[32:35], %[v_os_b2], s[20:23], 0 offen \n"
"buffer_load_dwordx4 acc[36:39], %[v_os_b2], s[20:23], 0 offen offset:1024 \n"
"buffer_load_dwordx4 acc[40:43], %[v_os_b2], s[20:23], 0 offen offset:2048 \n"
"buffer_load_dwordx4 acc[44:47], %[v_os_b2], s[20:23], 0 offen offset:3072 \n"
"buffer_load_dwordx4 acc[48:51], %[v_os_b3], s[20:23], 0 offen \n"
"buffer_load_dwordx4 acc[52:55], %[v_os_b3], s[20:23], 0 offen offset:1024 \n"
"buffer_load_dwordx4 acc[56:59], %[v_os_b3], s[20:23], 0 offen offset:2048 \n"
"buffer_load_dwordx4 acc[60:63], %[v_os_b3], s[20:23], 0 offen offset:3072 \n"
"buffer_load_dwordx4 acc[64:67], %[v_os_b4], s[20:23], 0 offen \n"
"buffer_load_dwordx4 acc[68:71], %[v_os_b4], s[20:23], 0 offen offset:1024 \n"
"buffer_load_dwordx4 acc[72:75], %[v_os_b4], s[20:23], 0 offen offset:2048 \n"
"buffer_load_dwordx4 acc[76:79], %[v_os_b4], s[20:23], 0 offen offset:3072 \n"
"buffer_load_dwordx4 acc[80:83], %[v_os_b5], s[20:23], 0 offen \n"
"buffer_load_dwordx4 acc[84:87], %[v_os_b5], s[20:23], 0 offen offset:1024 \n"
"buffer_load_dwordx4 acc[88:91], %[v_os_b5], s[20:23], 0 offen offset:2048 \n"
"buffer_load_dwordx4 acc[92:95], %[v_os_b5], s[20:23], 0 offen offset:3072 \n"
"buffer_load_dwordx4 acc[96:99], %[v_os_b6], s[20:23], 0 offen \n"
"buffer_load_dwordx4 acc[100:103], %[v_os_b6], s[20:23], 0 offen offset:1024 \n"
"buffer_load_dwordx4 acc[104:107], %[v_os_b6], s[20:23], 0 offen offset:2048 \n"
"buffer_load_dwordx4 acc[108:111], %[v_os_b6], s[20:23], 0 offen offset:3072 \n"
"buffer_load_dwordx4 acc[112:115], %[v_os_b7], s[20:23], 0 offen \n"
"buffer_load_dwordx4 acc[116:119], %[v_os_b7], s[20:23], 0 offen offset:1024 \n"
"buffer_load_dwordx4 acc[120:123], %[v_os_b7], s[20:23], 0 offen offset:2048 \n"
"buffer_load_dwordx4 acc[124:127], %[v_os_b7], s[20:23], 0 offen offset:3072 \n"
"s_cmp_gt_i32 %[s_loop_cnt] 1 ; move b with cond \n"
"s_cselect_b32 s86, %[s_tile_os_b], 0 ; move b with cond \n"
"s_add_u32 s20, s86, s20 ; move b with cond \n"
"s_addc_u32 s21, 0, s21 ; move b with cond \n"
"s_waitcnt vmcnt(40) \n"
"s_barrier \n"
"ds_read_b128 v[64:67], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_0]\n" // 1024: N stride, 64
// K stride
"ds_read_b128 v[68:71], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_1]\n"
"ds_read_b128 v[72:75], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_2]\n"
"ds_read_b128 v[76:79], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_3]\n"
"ds_read_b128 v[80:83], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_4]\n"
"ds_read_b128 v[84:87], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_5]\n"
"ds_read_b128 v[88:91], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_6]\n"
"ds_read_b128 v[92:95], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_7]\n"
"L_start%=: \n"
" s_waitcnt vmcnt(24) & lgkmcnt(0) \n"
" s_barrier \n" _UK_MFMA_
" %[v_acc_0], acc[0:1], v[64:65], %[v_acc_0] \n" _UK_MFMA_
" %[v_acc_0], acc[2:3], v[66:67], %[v_acc_0] \n"
" buffer_load_dwordx4 acc[128:131], %[v_os_b0], s[20:23], 0 offen \n" _UK_MFMA_
" %[v_acc_0], acc[4:5], v[68:69], %[v_acc_0] \n" _UK_MFMA_
" %[v_acc_0], acc[6:7], v[70:71], %[v_acc_0] \n"
" buffer_load_dword %[v_os_a0], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n" _UK_MFMA_
" %[v_acc_0], acc[8:9], v[72:73], %[v_acc_0] \n" _UK_MFMA_
" %[v_acc_0], acc[10:11], v[74:75], %[v_acc_0] \n"
" buffer_load_dwordx4 acc[132:135], %[v_os_b0], s[20:23], 0 offen offset:1024 \n" _UK_MFMA_
" %[v_acc_0], acc[12:13], v[76:77], %[v_acc_0] \n" _UK_MFMA_
" %[v_acc_0], acc[14:15], v[78:79], %[v_acc_0] \n"
" buffer_load_dword %[v_os_a1], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n" _UK_MFMA_
" %[v_acc_1], acc[0:1], v[80:81], %[v_acc_1] \n" _UK_MFMA_
" %[v_acc_1], acc[2:3], v[82:83], %[v_acc_1] \n"
" buffer_load_dwordx4 acc[136:139], %[v_os_b0], s[20:23], 0 offen offset:2048 \n" _UK_MFMA_
" %[v_acc_1], acc[4:5], v[84:85], %[v_acc_1] \n" _UK_MFMA_
" %[v_acc_1], acc[6:7], v[86:87], %[v_acc_1] \n"
" buffer_load_dword %[v_os_a2], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n" _UK_MFMA_
" %[v_acc_1], acc[8:9], v[88:89], %[v_acc_1] \n" _UK_MFMA_
" %[v_acc_1], acc[10:11], v[90:91], %[v_acc_1] \n"
" buffer_load_dwordx4 acc[140:143], %[v_os_b0], s[20:23], 0 offen offset:3072 \n" _UK_MFMA_
" %[v_acc_1], acc[12:13], v[92:93], %[v_acc_1] \n" _UK_MFMA_
" %[v_acc_1], acc[14:15], v[94:95], %[v_acc_1] \n"
" buffer_load_dword %[v_os_a3], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n" _UK_MFMA_
" %[v_acc_2], acc[16:17], v[64:65], %[v_acc_2] \n" _UK_MFMA_
" %[v_acc_2], acc[18:19], v[66:67], %[v_acc_2] \n"
" buffer_load_dwordx4 acc[144:147], %[v_os_b1], s[20:23], 0 offen \n" _UK_MFMA_
" %[v_acc_2], acc[20:21], v[68:69], %[v_acc_2] \n" _UK_MFMA_
" %[v_acc_2], acc[22:23], v[70:71], %[v_acc_2] \n"
" buffer_load_dword %[v_os_a4], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n" _UK_MFMA_
" %[v_acc_2], acc[24:25], v[72:73], %[v_acc_2] \n" _UK_MFMA_
" %[v_acc_2], acc[26:27], v[74:75], %[v_acc_2] \n"
" buffer_load_dwordx4 acc[148:151], %[v_os_b1], s[20:23], 0 offen offset:1024 \n" _UK_MFMA_
" %[v_acc_2], acc[28:29], v[76:77], %[v_acc_2] \n" _UK_MFMA_
" %[v_acc_2], acc[30:31], v[78:79], %[v_acc_2] \n"
" buffer_load_dword %[v_os_a5], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n" _UK_MFMA_
" %[v_acc_3], acc[16:17], v[80:81], %[v_acc_3] \n" _UK_MFMA_
" %[v_acc_3], acc[18:19], v[82:83], %[v_acc_3] \n"
" buffer_load_dwordx4 acc[152:155], %[v_os_b1], s[20:23], 0 offen offset:2048 \n" _UK_MFMA_
" %[v_acc_3], acc[20:21], v[84:85], %[v_acc_3] \n" _UK_MFMA_
" %[v_acc_3], acc[22:23], v[86:87], %[v_acc_3] \n"
" buffer_load_dword %[v_os_a6], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n" _UK_MFMA_
" %[v_acc_3], acc[24:25], v[88:89], %[v_acc_3] \n" _UK_MFMA_
" %[v_acc_3], acc[26:27], v[90:91], %[v_acc_3] \n"
" buffer_load_dwordx4 acc[156:159], %[v_os_b1], s[20:23], 0 offen offset:3072 \n" _UK_MFMA_
" %[v_acc_3], acc[28:29], v[92:93], %[v_acc_3] \n" _UK_MFMA_
" %[v_acc_3], acc[30:31], v[94:95], %[v_acc_3] \n"
" buffer_load_dword %[v_os_a7], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[smem_sz], %[s_m0_init] \n"
" s_waitcnt vmcnt(32) \n" _UK_MFMA_
" %[v_acc_4], acc[32:33], v[64:65], %[v_acc_4] \n" _UK_MFMA_
" %[v_acc_4], acc[34:35], v[66:67], %[v_acc_4] \n"
" buffer_load_dwordx4 acc[160:163], %[v_os_b2], s[20:23], 0 offen \n" _UK_MFMA_
" %[v_acc_4], acc[36:37], v[68:69], %[v_acc_4] \n" _UK_MFMA_
" %[v_acc_4], acc[38:39], v[70:71], %[v_acc_4] \n"
" ds_read_b128 v[96:99], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_0] "
"\n" _UK_MFMA_ " %[v_acc_4], acc[40:41], v[72:73], %[v_acc_4] \n" _UK_MFMA_
" %[v_acc_4], acc[42:43], v[74:75], %[v_acc_4] \n"
" buffer_load_dwordx4 acc[164:167], %[v_os_b2], s[20:23], 0 offen offset:1024 \n" _UK_MFMA_
" %[v_acc_4], acc[44:45], v[76:77], %[v_acc_4] \n" _UK_MFMA_
" %[v_acc_4], acc[46:47], v[78:79], %[v_acc_4] \n"
" ds_read_b128 v[100:103], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_1] "
"\n" _UK_MFMA_ " %[v_acc_5], acc[32:33], v[80:81], %[v_acc_5] \n" _UK_MFMA_
" %[v_acc_5], acc[34:35], v[82:83], %[v_acc_5] \n"
" buffer_load_dwordx4 acc[168:171], %[v_os_b2], s[20:23], 0 offen offset:2048 \n" _UK_MFMA_
" %[v_acc_5], acc[36:37], v[84:85], %[v_acc_5] \n" _UK_MFMA_
" %[v_acc_5], acc[38:39], v[86:87], %[v_acc_5] \n"
" ds_read_b128 v[104:107], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_2] "
"\n" _UK_MFMA_ " %[v_acc_5], acc[40:41], v[88:89], %[v_acc_5] \n" _UK_MFMA_
" %[v_acc_5], acc[42:43], v[90:91], %[v_acc_5] \n"
" buffer_load_dwordx4 acc[172:175], %[v_os_b2], s[20:23], 0 offen offset:3072 \n" _UK_MFMA_
" %[v_acc_5], acc[44:45], v[92:93], %[v_acc_5] \n" _UK_MFMA_
" %[v_acc_5], acc[46:47], v[94:95], %[v_acc_5] \n"
" ds_read_b128 v[108:111], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_3] "
"\n" _UK_MFMA_ " %[v_acc_6], acc[48:49], v[64:65], %[v_acc_6] \n" _UK_MFMA_
" %[v_acc_6], acc[50:51], v[66:67], %[v_acc_6] \n"
" buffer_load_dwordx4 acc[176:179], %[v_os_b3], s[20:23], 0 offen \n" _UK_MFMA_
" %[v_acc_6], acc[52:53], v[68:69], %[v_acc_6] \n" _UK_MFMA_
" %[v_acc_6], acc[54:55], v[70:71], %[v_acc_6] \n"
" ds_read_b128 v[112:115], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_4] "
"\n" _UK_MFMA_ " %[v_acc_6], acc[56:57], v[72:73], %[v_acc_6] \n" _UK_MFMA_
" %[v_acc_6], acc[58:59], v[74:75], %[v_acc_6] \n"
" buffer_load_dwordx4 acc[180:183], %[v_os_b3], s[20:23], 0 offen offset:1024 \n" _UK_MFMA_
" %[v_acc_6], acc[60:61], v[76:77], %[v_acc_6] \n" _UK_MFMA_
" %[v_acc_6], acc[62:63], v[78:79], %[v_acc_6] \n"
" ds_read_b128 v[116:119], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_5] "
"\n" _UK_MFMA_ " %[v_acc_7], acc[48:49], v[80:81], %[v_acc_7] \n" _UK_MFMA_
" %[v_acc_7], acc[50:51], v[82:83], %[v_acc_7] \n"
" buffer_load_dwordx4 acc[184:187], %[v_os_b3], s[20:23], 0 offen offset:2048 \n" _UK_MFMA_
" %[v_acc_7], acc[52:53], v[84:85], %[v_acc_7] \n" _UK_MFMA_
" %[v_acc_7], acc[54:55], v[86:87], %[v_acc_7] \n"
" ds_read_b128 v[120:123], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_6] "
"\n" _UK_MFMA_ " %[v_acc_7], acc[56:57], v[88:89], %[v_acc_7] \n" _UK_MFMA_
" %[v_acc_7], acc[58:59], v[90:91], %[v_acc_7] \n"
" buffer_load_dwordx4 acc[188:191], %[v_os_b3], s[20:23], 0 offen offset:3072 \n" _UK_MFMA_
" %[v_acc_7], acc[60:61], v[92:93], %[v_acc_7] \n" _UK_MFMA_
" %[v_acc_7], acc[62:63], v[94:95], %[v_acc_7] \n"
" ds_read_b128 v[124:127], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_7] \n"
" s_waitcnt vmcnt(32) \n" _UK_MFMA_
" %[v_acc_8], acc[64:65], v[64:65], %[v_acc_8] \n" _UK_MFMA_
" %[v_acc_8], acc[66:67], v[66:67], %[v_acc_8] \n"
" buffer_load_dwordx4 acc[192:195], %[v_os_b4], s[20:23], 0 offen \n" _UK_MFMA_
" %[v_acc_8], acc[68:69], v[68:69], %[v_acc_8] \n" _UK_MFMA_
" %[v_acc_8], acc[70:71], v[70:71], %[v_acc_8] \n" _UK_MFMA_
" %[v_acc_8], acc[72:73], v[72:73], %[v_acc_8] \n" _UK_MFMA_
" %[v_acc_8], acc[74:75], v[74:75], %[v_acc_8] \n"
" buffer_load_dwordx4 acc[196:199], %[v_os_b4], s[20:23], 0 offen offset:1024 \n" _UK_MFMA_
" %[v_acc_8], acc[76:77], v[76:77], %[v_acc_8] \n" _UK_MFMA_
" %[v_acc_8], acc[78:79], v[78:79], %[v_acc_8] \n" _UK_MFMA_
" %[v_acc_9], acc[64:65], v[80:81], %[v_acc_9] \n" _UK_MFMA_
" %[v_acc_9], acc[66:67], v[82:83], %[v_acc_9] \n"
" buffer_load_dwordx4 acc[200:203], %[v_os_b4], s[20:23], 0 offen offset:2048 \n" _UK_MFMA_
" %[v_acc_9], acc[68:69], v[84:85], %[v_acc_9] \n" _UK_MFMA_
" %[v_acc_9], acc[70:71], v[86:87], %[v_acc_9] \n" _UK_MFMA_
" %[v_acc_9], acc[72:73], v[88:89], %[v_acc_9] \n" _UK_MFMA_
" %[v_acc_9], acc[74:75], v[90:91], %[v_acc_9] \n"
" buffer_load_dwordx4 acc[204:207], %[v_os_b4], s[20:23], 0 offen offset:3072 \n" _UK_MFMA_
" %[v_acc_9], acc[76:77], v[92:93], %[v_acc_9] \n" _UK_MFMA_
" %[v_acc_9], acc[78:79], v[94:95], %[v_acc_9] \n" _UK_MFMA_
" %[v_acc_10], acc[80:81], v[64:65], %[v_acc_10] \n" _UK_MFMA_
" %[v_acc_10], acc[82:83], v[66:67], %[v_acc_10] \n"
" buffer_load_dwordx4 acc[208:211], %[v_os_b5], s[20:23], 0 offen \n" _UK_MFMA_
" %[v_acc_10], acc[84:85], v[68:69], %[v_acc_10] \n" _UK_MFMA_
" %[v_acc_10], acc[86:87], v[70:71], %[v_acc_10] \n" _UK_MFMA_
" %[v_acc_10], acc[88:89], v[72:73], %[v_acc_10] \n" _UK_MFMA_
" %[v_acc_10], acc[90:91], v[74:75], %[v_acc_10] \n"
" buffer_load_dwordx4 acc[212:215], %[v_os_b5], s[20:23], 0 offen offset:1024 \n" _UK_MFMA_
" %[v_acc_10], acc[92:93], v[76:77], %[v_acc_10] \n" _UK_MFMA_
" %[v_acc_10], acc[94:95], v[78:79], %[v_acc_10] \n" _UK_MFMA_
" %[v_acc_11], acc[80:81], v[80:81], %[v_acc_11] \n" _UK_MFMA_
" %[v_acc_11], acc[82:83], v[82:83], %[v_acc_11] \n"
" buffer_load_dwordx4 acc[216:219], %[v_os_b5], s[20:23], 0 offen offset:2048 \n" _UK_MFMA_
" %[v_acc_11], acc[84:85], v[84:85], %[v_acc_11] \n" _UK_MFMA_
" %[v_acc_11], acc[86:87], v[86:87], %[v_acc_11] \n" _UK_MFMA_
" %[v_acc_11], acc[88:89], v[88:89], %[v_acc_11] \n" _UK_MFMA_
" %[v_acc_11], acc[90:91], v[90:91], %[v_acc_11] \n"
" buffer_load_dwordx4 acc[220:223], %[v_os_b5], s[20:23], 0 offen offset:3072 \n" _UK_MFMA_
" %[v_acc_11], acc[92:93], v[92:93], %[v_acc_11] \n" _UK_MFMA_
" %[v_acc_11], acc[94:95], v[94:95], %[v_acc_11] \n"
" s_waitcnt vmcnt(32) \n" _UK_MFMA_
" %[v_acc_12], acc[96:97], v[64:65], %[v_acc_12] \n" _UK_MFMA_
" %[v_acc_12], acc[98:99], v[66:67], %[v_acc_12] \n"
" buffer_load_dwordx4 acc[224:227], %[v_os_b6], s[20:23], 0 offen \n" _UK_MFMA_
" %[v_acc_12], acc[100:101], v[68:69], %[v_acc_12] \n" _UK_MFMA_
" %[v_acc_12], acc[102:103], v[70:71], %[v_acc_12] \n" _UK_MFMA_
" %[v_acc_12], acc[104:105], v[72:73], %[v_acc_12] \n" _UK_MFMA_
" %[v_acc_12], acc[106:107], v[74:75], %[v_acc_12] \n"
" buffer_load_dwordx4 acc[228:231], %[v_os_b6], s[20:23], 0 offen offset:1024 \n" _UK_MFMA_
" %[v_acc_12], acc[108:109], v[76:77], %[v_acc_12] \n" _UK_MFMA_
" %[v_acc_12], acc[110:111], v[78:79], %[v_acc_12] \n" _UK_MFMA_
" %[v_acc_13], acc[96:97], v[80:81], %[v_acc_13] \n" _UK_MFMA_
" %[v_acc_13], acc[98:99], v[82:83], %[v_acc_13] \n"
" buffer_load_dwordx4 acc[232:235], %[v_os_b6], s[20:23], 0 offen offset:2048 \n" _UK_MFMA_
" %[v_acc_13], acc[100:101], v[84:85], %[v_acc_13] \n" _UK_MFMA_
" %[v_acc_13], acc[102:103], v[86:87], %[v_acc_13] \n" _UK_MFMA_
" %[v_acc_13], acc[104:105], v[88:89], %[v_acc_13] \n" _UK_MFMA_
" %[v_acc_13], acc[106:107], v[90:91], %[v_acc_13] \n"
" buffer_load_dwordx4 acc[236:239], %[v_os_b6], s[20:23], 0 offen offset:3072 \n" _UK_MFMA_
" %[v_acc_13], acc[108:109], v[92:93], %[v_acc_13] \n" _UK_MFMA_
" %[v_acc_13], acc[110:111], v[94:95], %[v_acc_13] \n" _UK_MFMA_
" %[v_acc_14], acc[112:113], v[64:65], %[v_acc_14] \n" _UK_MFMA_
" %[v_acc_14], acc[114:115], v[66:67], %[v_acc_14] \n"
" buffer_load_dwordx4 acc[240:243], %[v_os_b7], s[20:23], 0 offen \n" _UK_MFMA_
" %[v_acc_14], acc[116:117], v[68:69], %[v_acc_14] \n" _UK_MFMA_
" %[v_acc_14], acc[118:119], v[70:71], %[v_acc_14] \n" _UK_MFMA_
" %[v_acc_14], acc[120:121], v[72:73], %[v_acc_14] \n" _UK_MFMA_
" %[v_acc_14], acc[122:123], v[74:75], %[v_acc_14] \n"
" buffer_load_dwordx4 acc[244:247], %[v_os_b7], s[20:23], 0 offen offset:1024 \n" _UK_MFMA_
" %[v_acc_14], acc[124:125], v[76:77], %[v_acc_14] \n" _UK_MFMA_
" %[v_acc_14], acc[126:127], v[78:79], %[v_acc_14] \n" _UK_MFMA_
" %[v_acc_15], acc[112:113], v[80:81], %[v_acc_15] \n" _UK_MFMA_
" %[v_acc_15], acc[114:115], v[82:83], %[v_acc_15] \n"
" buffer_load_dwordx4 acc[248:251], %[v_os_b7], s[20:23], 0 offen offset:2048 \n" _UK_MFMA_
" %[v_acc_15], acc[116:117], v[84:85], %[v_acc_15] \n" _UK_MFMA_
" %[v_acc_15], acc[118:119], v[86:87], %[v_acc_15] \n" _UK_MFMA_
" %[v_acc_15], acc[120:121], v[88:89], %[v_acc_15] \n" _UK_MFMA_
" %[v_acc_15], acc[122:123], v[90:91], %[v_acc_15] \n"
" buffer_load_dwordx4 acc[252:255], %[v_os_b7], s[20:23], 0 offen offset:3072\n" _UK_MFMA_
" %[v_acc_15], acc[124:125], v[92:93], %[v_acc_15] \n" _UK_MFMA_
" %[v_acc_15], acc[126:127], v[94:95], %[v_acc_15] \n"
" s_sub_i32 %[s_loop_cnt], %[s_loop_cnt], 1 \n"
" s_cmp_gt_i32 %[s_loop_cnt] 0 \n"
" s_cbranch_scc0 L_end%= \n"
" s_cmp_gt_i32 %[s_loop_cnt] 2 ; move a with cond \n"
" s_cselect_b32 s86, %[s_tile_os_a], 0 \n"
" s_add_u32 s16, s86, s16 \n"
" s_addc_u32 s17, 0, s17 \n"
" s_cmp_gt_i32 %[s_loop_cnt] 1 ; move b with cond \n"
" s_cselect_b32 s86, %[s_tile_os_b], 0 \n"
" s_add_u32 s20, s86, s20 \n"
" s_addc_u32 s21, 0, s21 \n"
" ;------------------------------------------ \n"
" s_waitcnt vmcnt(24) & lgkmcnt(0) \n"
" s_barrier \n" _UK_MFMA_
" %[v_acc_0], acc[128:129], v[96:97], %[v_acc_0] \n" _UK_MFMA_
" %[v_acc_0], acc[130:131], v[98:99], %[v_acc_0] \n"
" buffer_load_dwordx4 acc[0:3], %[v_os_b0], s[20:23], 0 offen \n" _UK_MFMA_
" %[v_acc_0], acc[132:133], v[100:101], %[v_acc_0] \n" _UK_MFMA_
" %[v_acc_0], acc[134:135], v[102:103], %[v_acc_0] \n"
" buffer_load_dword %[v_os_a0], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n" _UK_MFMA_
" %[v_acc_0], acc[136:137], v[104:105], %[v_acc_0] \n" _UK_MFMA_
" %[v_acc_0], acc[138:139], v[106:107], %[v_acc_0] \n"
" buffer_load_dwordx4 acc[4:7], %[v_os_b0], s[20:23], 0 offen offset:1024 \n" _UK_MFMA_
" %[v_acc_0], acc[140:141], v[108:109], %[v_acc_0] \n" _UK_MFMA_
" %[v_acc_0], acc[142:143], v[110:111], %[v_acc_0] \n"
" buffer_load_dword %[v_os_a1], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n" _UK_MFMA_
" %[v_acc_1], acc[128:129], v[112:113], %[v_acc_1] \n" _UK_MFMA_
" %[v_acc_1], acc[130:131], v[114:115], %[v_acc_1] \n"
" buffer_load_dwordx4 acc[8:11], %[v_os_b0], s[20:23], 0 offen offset:2048 \n" _UK_MFMA_
" %[v_acc_1], acc[132:133], v[116:117], %[v_acc_1] \n" _UK_MFMA_
" %[v_acc_1], acc[134:135], v[118:119], %[v_acc_1] \n"
" buffer_load_dword %[v_os_a2], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n" _UK_MFMA_
" %[v_acc_1], acc[136:137], v[120:121], %[v_acc_1] \n" _UK_MFMA_
" %[v_acc_1], acc[138:139], v[122:123], %[v_acc_1] \n"
" buffer_load_dwordx4 acc[12:15], %[v_os_b0], s[20:23], 0 offen offset:3072 \n" _UK_MFMA_
" %[v_acc_1], acc[140:141], v[124:125], %[v_acc_1] \n" _UK_MFMA_
" %[v_acc_1], acc[142:143], v[126:127], %[v_acc_1] \n"
" buffer_load_dword %[v_os_a3], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n" _UK_MFMA_
" %[v_acc_2], acc[144:145], v[96:97], %[v_acc_2] \n" _UK_MFMA_
" %[v_acc_2], acc[146:147], v[98:99], %[v_acc_2] \n"
" buffer_load_dwordx4 acc[16:19], %[v_os_b1], s[20:23], 0 offen \n" _UK_MFMA_
" %[v_acc_2], acc[148:149], v[100:101], %[v_acc_2] \n" _UK_MFMA_
" %[v_acc_2], acc[150:151], v[102:103], %[v_acc_2] \n"
" buffer_load_dword %[v_os_a4], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n" _UK_MFMA_
" %[v_acc_2], acc[152:153], v[104:105], %[v_acc_2] \n" _UK_MFMA_
" %[v_acc_2], acc[154:155], v[106:107], %[v_acc_2] \n"
" buffer_load_dwordx4 acc[20:23], %[v_os_b1], s[20:23], 0 offen offset:1024 \n" _UK_MFMA_
" %[v_acc_2], acc[156:157], v[108:109], %[v_acc_2] \n" _UK_MFMA_
" %[v_acc_2], acc[158:159], v[110:111], %[v_acc_2] \n"
" buffer_load_dword %[v_os_a5], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n" _UK_MFMA_
" %[v_acc_3], acc[144:145], v[112:113], %[v_acc_3] \n" _UK_MFMA_
" %[v_acc_3], acc[146:147], v[114:115], %[v_acc_3] \n"
" buffer_load_dwordx4 acc[24:27], %[v_os_b1], s[20:23], 0 offen offset:2048 \n" _UK_MFMA_
" %[v_acc_3], acc[148:149], v[116:117], %[v_acc_3] \n" _UK_MFMA_
" %[v_acc_3], acc[150:151], v[118:119], %[v_acc_3] \n"
" buffer_load_dword %[v_os_a6], s[16:19], 0 offen lds \n"
" s_add_u32 m0, %[s_size_per_issue], m0 \n" _UK_MFMA_
" %[v_acc_3], acc[152:153], v[120:121], %[v_acc_3] \n" _UK_MFMA_
" %[v_acc_3], acc[154:155], v[122:123], %[v_acc_3] \n"
" buffer_load_dwordx4 acc[28:31], %[v_os_b1], s[20:23], 0 offen offset:3072 \n" _UK_MFMA_
" %[v_acc_3], acc[156:157], v[124:125], %[v_acc_3] \n" _UK_MFMA_
" %[v_acc_3], acc[158:159], v[126:127], %[v_acc_3] \n"
" buffer_load_dword %[v_os_a7], s[16:19], 0 offen lds \n"
" s_add_u32 m0, 0, %[s_m0_init] \n"
" s_waitcnt vmcnt(32) \n" _UK_MFMA_
" %[v_acc_4], acc[160:161], v[96:97], %[v_acc_4] \n" _UK_MFMA_
" %[v_acc_4], acc[162:163], v[98:99], %[v_acc_4] \n"
" buffer_load_dwordx4 acc[32:35], %[v_os_b2], s[20:23], 0 offen \n" _UK_MFMA_
" %[v_acc_4], acc[164:165], v[100:101], %[v_acc_4] \n" _UK_MFMA_
" %[v_acc_4], acc[166:167], v[102:103], %[v_acc_4] \n"
" ds_read_b128 v[64:67], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_0] \n" _UK_MFMA_
" %[v_acc_4], acc[168:169], v[104:105], %[v_acc_4] \n" _UK_MFMA_
" %[v_acc_4], acc[170:171], v[106:107], %[v_acc_4] \n"
" buffer_load_dwordx4 acc[36:39], %[v_os_b2], s[20:23], 0 offen offset:1024 \n" _UK_MFMA_
" %[v_acc_4], acc[172:173], v[108:109], %[v_acc_4] \n" _UK_MFMA_
" %[v_acc_4], acc[174:175], v[110:111], %[v_acc_4] \n"
" ds_read_b128 v[68:71], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_1] \n" _UK_MFMA_
" %[v_acc_5], acc[160:161], v[112:113], %[v_acc_5] \n" _UK_MFMA_
" %[v_acc_5], acc[162:163], v[114:115], %[v_acc_5] \n"
" buffer_load_dwordx4 acc[40:43], %[v_os_b2], s[20:23], 0 offen offset:2048 \n" _UK_MFMA_
" %[v_acc_5], acc[164:165], v[116:117], %[v_acc_5] \n" _UK_MFMA_
" %[v_acc_5], acc[166:167], v[118:119], %[v_acc_5] \n"
" ds_read_b128 v[72:75], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_2] "
"\n" _UK_MFMA_ " %[v_acc_5], acc[168:169], v[120:121], %[v_acc_5] \n" _UK_MFMA_
" %[v_acc_5], acc[170:171], v[122:123], %[v_acc_5] \n"
" buffer_load_dwordx4 acc[44:47], %[v_os_b2], s[20:23], 0 offen offset:3072 \n" _UK_MFMA_
" %[v_acc_5], acc[172:173], v[124:125], %[v_acc_5] \n" _UK_MFMA_
" %[v_acc_5], acc[174:175], v[126:127], %[v_acc_5] \n"
" ds_read_b128 v[76:79], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_3] "
"\n" _UK_MFMA_ " %[v_acc_6], acc[176:177], v[96:97], %[v_acc_6] \n" _UK_MFMA_
" %[v_acc_6], acc[178:179], v[98:99], %[v_acc_6] \n"
" buffer_load_dwordx4 acc[48:51], %[v_os_b3], s[20:23], 0 offen \n" _UK_MFMA_
" %[v_acc_6], acc[180:181], v[100:101], %[v_acc_6] \n" _UK_MFMA_
" %[v_acc_6], acc[182:183], v[102:103], %[v_acc_6] \n"
" ds_read_b128 v[80:83], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_4] "
"\n" _UK_MFMA_ " %[v_acc_6], acc[184:185], v[104:105], %[v_acc_6] \n" _UK_MFMA_
" %[v_acc_6], acc[186:187], v[106:107], %[v_acc_6] \n"
" buffer_load_dwordx4 acc[52:55], %[v_os_b3], s[20:23], 0 offen offset:1024 \n" _UK_MFMA_
" %[v_acc_6], acc[188:189], v[108:109], %[v_acc_6] \n" _UK_MFMA_
" %[v_acc_6], acc[190:191], v[110:111], %[v_acc_6] \n"
" ds_read_b128 v[84:87], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_5] "
"\n" _UK_MFMA_ " %[v_acc_7], acc[176:177], v[112:113], %[v_acc_7] \n" _UK_MFMA_
" %[v_acc_7], acc[178:179], v[114:115], %[v_acc_7] \n"
" buffer_load_dwordx4 acc[56:59], %[v_os_b3], s[20:23], 0 offen offset:2048 \n" _UK_MFMA_
" %[v_acc_7], acc[180:181], v[116:117], %[v_acc_7] \n" _UK_MFMA_
" %[v_acc_7], acc[182:183], v[118:119], %[v_acc_7] \n"
" ds_read_b128 v[88:91], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_6] "
"\n" _UK_MFMA_ " %[v_acc_7], acc[184:185], v[120:121], %[v_acc_7] \n" _UK_MFMA_
" %[v_acc_7], acc[186:187], v[122:123], %[v_acc_7] \n"
" buffer_load_dwordx4 acc[60:63], %[v_os_b3], s[20:23], 0 offen offset:3072 \n" _UK_MFMA_
" %[v_acc_7], acc[188:189], v[124:125], %[v_acc_7] \n" _UK_MFMA_
" %[v_acc_7], acc[190:191], v[126:127], %[v_acc_7] \n"
" ds_read_b128 v[92:95], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_7] \n"
" s_waitcnt vmcnt(32) \n" _UK_MFMA_
" %[v_acc_8], acc[192:193], v[96:97], %[v_acc_8] \n" _UK_MFMA_
" %[v_acc_8], acc[194:195], v[98:99], %[v_acc_8] \n"
" buffer_load_dwordx4 acc[64:67], %[v_os_b4], s[20:23], 0 offen \n" _UK_MFMA_
" %[v_acc_8], acc[196:197], v[100:101], %[v_acc_8] \n" _UK_MFMA_
" %[v_acc_8], acc[198:199], v[102:103], %[v_acc_8] \n" _UK_MFMA_
" %[v_acc_8], acc[200:201], v[104:105], %[v_acc_8] \n" _UK_MFMA_
" %[v_acc_8], acc[202:203], v[106:107], %[v_acc_8] \n"
" buffer_load_dwordx4 acc[68:71], %[v_os_b4], s[20:23], 0 offen offset:1024 \n" _UK_MFMA_
" %[v_acc_8], acc[204:205], v[108:109], %[v_acc_8] \n" _UK_MFMA_
" %[v_acc_8], acc[206:207], v[110:111], %[v_acc_8] \n" _UK_MFMA_
" %[v_acc_9], acc[192:193], v[112:113], %[v_acc_9] \n" _UK_MFMA_
" %[v_acc_9], acc[194:195], v[114:115], %[v_acc_9] \n"
" buffer_load_dwordx4 acc[72:75], %[v_os_b4], s[20:23], 0 offen offset:2048 \n" _UK_MFMA_
" %[v_acc_9], acc[196:197], v[116:117], %[v_acc_9] \n" _UK_MFMA_
" %[v_acc_9], acc[198:199], v[118:119], %[v_acc_9] \n" _UK_MFMA_
" %[v_acc_9], acc[200:201], v[120:121], %[v_acc_9] \n" _UK_MFMA_
" %[v_acc_9], acc[202:203], v[122:123], %[v_acc_9] \n"
" buffer_load_dwordx4 acc[76:79], %[v_os_b4], s[20:23], 0 offen offset:3072 \n" _UK_MFMA_
" %[v_acc_9], acc[204:205], v[124:125], %[v_acc_9] \n" _UK_MFMA_
" %[v_acc_9], acc[206:207], v[126:127], %[v_acc_9] \n" _UK_MFMA_
" %[v_acc_10], acc[208:209], v[96:97], %[v_acc_10] \n" _UK_MFMA_
" %[v_acc_10], acc[210:211], v[98:99], %[v_acc_10] \n"
" buffer_load_dwordx4 acc[80:83], %[v_os_b5], s[20:23], 0 offen \n" _UK_MFMA_
" %[v_acc_10], acc[212:213], v[100:101], %[v_acc_10] \n" _UK_MFMA_
" %[v_acc_10], acc[214:215], v[102:103], %[v_acc_10] \n" _UK_MFMA_
" %[v_acc_10], acc[216:217], v[104:105], %[v_acc_10] \n" _UK_MFMA_
" %[v_acc_10], acc[218:219], v[106:107], %[v_acc_10] \n"
" buffer_load_dwordx4 acc[84:87], %[v_os_b5], s[20:23], 0 offen offset:1024 \n" _UK_MFMA_
" %[v_acc_10], acc[220:221], v[108:109], %[v_acc_10] \n" _UK_MFMA_
" %[v_acc_10], acc[222:223], v[110:111], %[v_acc_10] \n" _UK_MFMA_
" %[v_acc_11], acc[208:209], v[112:113], %[v_acc_11] \n" _UK_MFMA_
" %[v_acc_11], acc[210:211], v[114:115], %[v_acc_11] \n"
" buffer_load_dwordx4 acc[88:91], %[v_os_b5], s[20:23], 0 offen offset:2048 \n" _UK_MFMA_
" %[v_acc_11], acc[212:213], v[116:117], %[v_acc_11] \n" _UK_MFMA_
" %[v_acc_11], acc[214:215], v[118:119], %[v_acc_11] \n" _UK_MFMA_
" %[v_acc_11], acc[216:217], v[120:121], %[v_acc_11] \n" _UK_MFMA_
" %[v_acc_11], acc[218:219], v[122:123], %[v_acc_11] \n"
" buffer_load_dwordx4 acc[92:95], %[v_os_b5], s[20:23], 0 offen offset:3072 \n" _UK_MFMA_
" %[v_acc_11], acc[220:221], v[124:125], %[v_acc_11] \n" _UK_MFMA_
" %[v_acc_11], acc[222:223], v[126:127], %[v_acc_11] \n"
" s_waitcnt vmcnt(32) \n" _UK_MFMA_
" %[v_acc_12], acc[224:225], v[96:97], %[v_acc_12] \n" _UK_MFMA_
" %[v_acc_12], acc[226:227], v[98:99], %[v_acc_12] \n"
" buffer_load_dwordx4 acc[96:99], %[v_os_b6], s[20:23], 0 offen \n" _UK_MFMA_
" %[v_acc_12], acc[228:229], v[100:101], %[v_acc_12] \n" _UK_MFMA_
" %[v_acc_12], acc[230:231], v[102:103], %[v_acc_12] \n" _UK_MFMA_
" %[v_acc_12], acc[232:233], v[104:105], %[v_acc_12] \n" _UK_MFMA_
" %[v_acc_12], acc[234:235], v[106:107], %[v_acc_12] \n"
" buffer_load_dwordx4 acc[100:103], %[v_os_b6], s[20:23], 0 offen offset:1024 \n" _UK_MFMA_
" %[v_acc_12], acc[236:237], v[108:109], %[v_acc_12] \n" _UK_MFMA_
" %[v_acc_12], acc[238:239], v[110:111], %[v_acc_12] \n" _UK_MFMA_
" %[v_acc_13], acc[224:225], v[112:113], %[v_acc_13] \n" _UK_MFMA_
" %[v_acc_13], acc[226:227], v[114:115], %[v_acc_13] \n"
" buffer_load_dwordx4 acc[104:107], %[v_os_b6], s[20:23], 0 offen offset:2048 \n" _UK_MFMA_
" %[v_acc_13], acc[228:229], v[116:117], %[v_acc_13] \n" _UK_MFMA_
" %[v_acc_13], acc[230:231], v[118:119], %[v_acc_13] \n" _UK_MFMA_
" %[v_acc_13], acc[232:233], v[120:121], %[v_acc_13] \n" _UK_MFMA_
" %[v_acc_13], acc[234:235], v[122:123], %[v_acc_13] \n"
" buffer_load_dwordx4 acc[108:111], %[v_os_b6], s[20:23], 0 offen offset:3072 \n" _UK_MFMA_
" %[v_acc_13], acc[236:237], v[124:125], %[v_acc_13] \n" _UK_MFMA_
" %[v_acc_13], acc[238:239], v[126:127], %[v_acc_13] \n" _UK_MFMA_
" %[v_acc_14], acc[240:241], v[96:97], %[v_acc_14] \n" _UK_MFMA_
" %[v_acc_14], acc[242:243], v[98:99], %[v_acc_14] \n"
" buffer_load_dwordx4 acc[112:115], %[v_os_b7], s[20:23], 0 offen \n" _UK_MFMA_
" %[v_acc_14], acc[244:245], v[100:101], %[v_acc_14] \n" _UK_MFMA_
" %[v_acc_14], acc[246:247], v[102:103], %[v_acc_14] \n" _UK_MFMA_
" %[v_acc_14], acc[248:249], v[104:105], %[v_acc_14] \n" _UK_MFMA_
" %[v_acc_14], acc[250:251], v[106:107], %[v_acc_14] \n"
" buffer_load_dwordx4 acc[116:119], %[v_os_b7], s[20:23], 0 offen offset:1024 \n" _UK_MFMA_
" %[v_acc_14], acc[252:253], v[108:109], %[v_acc_14] \n" _UK_MFMA_
" %[v_acc_14], acc[254:255], v[110:111], %[v_acc_14] \n" _UK_MFMA_
" %[v_acc_15], acc[240:241], v[112:113], %[v_acc_15] \n" _UK_MFMA_
" %[v_acc_15], acc[242:243], v[114:115], %[v_acc_15] \n"
" buffer_load_dwordx4 acc[120:123], %[v_os_b7], s[20:23], 0 offen offset:2048 \n" _UK_MFMA_
" %[v_acc_15], acc[244:245], v[116:117], %[v_acc_15] \n" _UK_MFMA_
" %[v_acc_15], acc[246:247], v[118:119], %[v_acc_15] \n" _UK_MFMA_
" %[v_acc_15], acc[248:249], v[120:121], %[v_acc_15] \n" _UK_MFMA_
" %[v_acc_15], acc[250:251], v[122:123], %[v_acc_15] \n"
" buffer_load_dwordx4 acc[124:127], %[v_os_b7], s[20:23], 0 offen offset:3072 \n" _UK_MFMA_
" %[v_acc_15], acc[252:253], v[124:125], %[v_acc_15] \n" _UK_MFMA_
" %[v_acc_15], acc[254:255], v[126:127], %[v_acc_15] \n"
" s_sub_i32 %[s_loop_cnt], %[s_loop_cnt], 1 \n"
" s_cmp_gt_i32 %[s_loop_cnt] 0 \n"
" s_cbranch_scc0 L_end%= \n"
" s_cmp_gt_i32 %[s_loop_cnt] 2 ; move a with cond \n"
" s_cselect_b32 s86, %[s_tile_os_a], 0 \n"
" s_add_u32 s16, s86, s16 \n"
" s_addc_u32 s17, 0, s17 \n"
" s_cmp_gt_i32 %[s_loop_cnt] 1 ; move b with cond \n"
" s_cselect_b32 s86, %[s_tile_os_b], 0 \n"
" s_add_u32 s20, s86, s20 \n"
" s_addc_u32 s21, 0, s21 \n"
" s_branch L_start%= \n"
"L_end%=: \n"
" s_nop 2 \n"
#undef _UK_MFMA_
include/ck_tile/ops/fmha.hpp
View file @ 2a768425
......@@ -29,6 +29,8 @@
#include "ck_tile/ops/fmha/pipeline/block_fmha_fwd_appendkv_pipeline_default_policy.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_combine_pipeline.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_combine_pipeline_default_policy.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_nwarp_sshuffle_qr_ks_vs.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_nwarp_sshuffle_qr_ks_vs_default_policy.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_qr_ks_vs.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_qr_ks_vs_default_policy.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_pipeline_enum.hpp"
......
include/ck_tile/ops/fmha/kernel/fmha_fwd_kernel.hpp
View file @ 2a768425
......@@ -71,7 +71,8 @@ struct FmhaFwdKernel
using bfs = typename FmhaPipeline::BlockFmhaShape;
using g0br = typename bfs::Gemm0BlockWarps;
using g1br = typename bfs::Gemm1BlockWarps;
using gwt = typename bfs::Gemm0WarpTile;
using g0wt = typename bfs::Gemm0WarpTile;
using g1wt = typename bfs::Gemm1WarpTile;
#define _SS_ std::string
#define _TS_ std::to_string
auto pn = [&] () {
......@@ -88,7 +89,8 @@ struct FmhaFwdKernel
_TS_(bfs::kN1) + "x" + _TS_(bfs::kK1) + "x" + _TS_(bfs::kQKHeaddim) + "_" +
"r" + _TS_(g0br::at(ck_tile::number<0>{})) + "x" + _TS_(g0br::at(ck_tile::number<1>{})) + "x" + _TS_(g0br::at(ck_tile::number<2>{})) + "_" +
"r" + _TS_(g1br::at(ck_tile::number<0>{})) + "x" + _TS_(g1br::at(ck_tile::number<1>{})) + "x" + _TS_(g1br::at(ck_tile::number<2>{})) + "_" +
"w" + _TS_(gwt::at(ck_tile::number<0>{})) + "x" + _TS_(gwt::at(ck_tile::number<1>{})) + "x" + _TS_(gwt::at(ck_tile::number<2>{})) + "_" +
"w" + _TS_(g0wt::at(ck_tile::number<0>{})) + "x" + _TS_(g0wt::at(ck_tile::number<1>{})) + "x" + _TS_(g0wt::at(ck_tile::number<2>{})) + "_" +
"w" + _TS_(g1wt::at(ck_tile::number<0>{})) + "x" + _TS_(g1wt::at(ck_tile::number<1>{})) + "x" + _TS_(g1wt::at(ck_tile::number<2>{})) + "_" +
(kBlockPerCuInput == -1 ? "" : ("o" + _TS_(kBlockPerCu) + "_")) + _SS_(FmhaPipeline::name) + "_" +
"v" + (std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor> ? "r" : "c") + (pn.empty() ? "" : "_" + pn) +
(BiasEnum == BlockAttentionBiasEnum::NO_BIAS ? _SS_("") : (_SS_("_") + BlockAttentionBiasEnumToStr<BiasEnum>::name)) +
......
include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_combine_kernel.hpp
View file @ 2a768425
......@@ -8,9 +8,11 @@ namespace ck_tile {
template <typename TilePartitioner_, typename FmhaPipeline_, typename EpiloguePipeline_>
struct FmhaFwdSplitKVCombineKernel
{
using TilePartitioner = remove_cvref_t<TilePartitioner_>;
using FmhaPipeline = remove_cvref_t<FmhaPipeline_>;
using EpiloguePipeline = remove_cvref_t<EpiloguePipeline_>;
using TilePartitioner = remove_cvref_t<TilePartitioner_>;
using FmhaPipeline = remove_cvref_t<FmhaPipeline_>;
using EpiloguePipeline = remove_cvref_t<EpiloguePipeline_>;
static constexpr index_t kNumWarps = FmhaPipeline::kNumWarps;
static constexpr index_t kBlockSize = FmhaPipeline::kBlockSize;
static constexpr index_t kBlockPerCu = FmhaPipeline::kBlockPerCu;
static_assert(kBlockPerCu > 0);
......@@ -50,8 +52,7 @@ struct FmhaFwdSplitKVCombineKernel
return
_SS_("fmha_fwd_splitkv_combine_d") + _TS_(FmhaPipeline::kHeadDimV) + "_" + _SS_(t2s<ODataType>::name) +
"_" + (kIsGroupMode ? "group" : "batch") + "_"
"b" + _TS_(FmhaPipeline::kM0) + "x" +
_TS_(FmhaPipeline::kN1) + "_" +
"b" + _TS_(FmhaPipeline::kN1) + "_" +
(kBlockPerCuInput == -1 ? "" : ("o" + _TS_(kBlockPerCu) + "_")) +
_SS_(FmhaPipeline::name) +
(pn.empty() ? "" : "_" + pn) +
......@@ -339,37 +340,56 @@ struct FmhaFwdSplitKVCombineKernel
number<FmhaPipeline::kAlignmentOacc>{},
number<1>{});
// read 4 * (kM0, kN1) o_acc tiles simultaneously by 4 warps
const auto o_acc_dram_view = pad_tensor_view(
o_acc_dram_naive,
make_tuple(number<1>{}, number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kN1>{}),
sequence<false, kPadSeqLenQ, kPadHeadDimV>{});
make_tuple(
number<kNumWarps>{}, number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kN1>{}),
sequence<true, kPadSeqLenQ, kPadHeadDimV>{});
const index_t padded_num_splits =
o_acc_dram_view.get_tensor_descriptor().get_lengths()[number<0>{}];
const index_t padded_seqlen_q =
o_acc_dram_view.get_tensor_descriptor().get_lengths()[number<1>{}];
const index_t padded_hdim_v =
o_acc_dram_view.get_tensor_descriptor().get_lengths()[number<2>{}];
return transform_tensor_view(
const index_t num_m_tiles = integer_divide_floor(padded_seqlen_q, FmhaPipeline::kM0);
// transform tensor view by following steps, given shape: (padded_num_splits,
// padded_seqlen_q, padded_hdim_v)
// 1. unmerge to (padded_num_splits, num_m_tiles, kM0, padded_hdim_v)
// 2. transpose to (num_m_tiles, padded_num_splits, kM0, padded_hdim_v)
// 3. merge to (num_m_tiles * padded_num_splits * kM0, padded_hdim_v)
auto transposed = transform_tensor_view(
o_acc_dram_view,
make_tuple(make_merge_transform(make_tuple(kargs.num_splits, padded_seqlen_q)),
make_tuple(make_pass_through_transform(padded_num_splits),
make_unmerge_transform(make_tuple(num_m_tiles, FmhaPipeline::kM0)),
make_pass_through_transform(padded_hdim_v)),
make_tuple(sequence<0, 1>{}, sequence<2>{}),
make_tuple(sequence<0>{}, sequence<1>{}, sequence<2>{}),
make_tuple(sequence<1>{}, sequence<0, 2>{}, sequence<3>{}));
return transform_tensor_view(
transposed,
make_tuple(make_merge_transform(
make_tuple(num_m_tiles, padded_num_splits, FmhaPipeline::kM0)),
make_pass_through_transform(padded_hdim_v)),
make_tuple(sequence<0, 1, 2>{}, sequence<3>{}),
make_tuple(sequence<0>{}, sequence<1>{}));
}();
auto lse_acc_dram_window = make_tile_window(
lse_acc_dram,
[&]() {
return make_tuple(number<FmhaPipeline::kMaxSplits>{}, number<FmhaPipeline::kM0>{});
}(),
make_tuple(number<FmhaPipeline::kMaxSplits>{}, number<FmhaPipeline::kM0>{}),
{0, i_m0});
const index_t padded_num_splits =
integer_divide_ceil(kargs.num_splits, kNumWarps) * kNumWarps;
auto o_acc_dram_window = make_tile_window(
o_acc_dram,
[&]() {
return make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kN1>{});
}(),
{i_m0, i_n1});
make_tuple(number<kNumWarps * FmhaPipeline::kM0>{}, number<FmhaPipeline::kN1>{}),
{i_tile_m * padded_num_splits * FmhaPipeline::kM0, i_n1});
// LSE DRAM window
auto lse_dram_window = [&, i_nhead_ = i_nhead]() {
......@@ -410,7 +430,6 @@ struct FmhaFwdSplitKVCombineKernel
identity{}, // lse_element_func
composes(saturates<fp8_t>{}, scales{kargs.scale_o}), // o_acc_element_func
kargs.num_splits,
kargs.seqlen_q,
smem_ptr);
}
else
......@@ -419,7 +438,6 @@ struct FmhaFwdSplitKVCombineKernel
o_acc_dram_window,
lse_dram_window,
kargs.num_splits,
kargs.seqlen_q,
smem_ptr);
}
}();
......
include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_kernel.hpp
View file @ 2a768425
......@@ -45,6 +45,7 @@ struct FmhaFwdSplitKVKernel
static constexpr bool kPadHeadDimQ = FmhaPipeline::kPadHeadDimQ;
static constexpr bool kPadHeadDimV = FmhaPipeline::kPadHeadDimV;
static constexpr auto BiasEnum = FmhaPipeline::BiasEnum;
static constexpr bool kStoreLSE = FmhaPipeline::kStoreLSE;
static constexpr bool kDoFp8StaticQuant = FmhaPipeline::Problem::kDoFp8StaticQuant;
static constexpr bool kIsPagedKV = FmhaPipeline::Problem::kIsPagedKV;
......@@ -67,7 +68,8 @@ struct FmhaFwdSplitKVKernel
using bfs = typename FmhaPipeline::BlockFmhaShape;
using g0br = typename bfs::Gemm0BlockWarps;
using g1br = typename bfs::Gemm1BlockWarps;
using gwt = typename bfs::Gemm0WarpTile;
using g0wt = typename bfs::Gemm0WarpTile;
using g1wt = typename bfs::Gemm1WarpTile;
#define _SS_ std::string
#define _TS_ std::to_string
auto pn = [&] () {
......@@ -84,11 +86,12 @@ struct FmhaFwdSplitKVKernel
_TS_(bfs::kN1) + "x" + _TS_(bfs::kK1) + "x" + _TS_(bfs::kQKHeaddim) + "_" +
"r" + _TS_(g0br::at(ck_tile::number<0>{})) + "x" + _TS_(g0br::at(ck_tile::number<1>{})) + "x" + _TS_(g0br::at(ck_tile::number<2>{})) + "_" +
"r" + _TS_(g1br::at(ck_tile::number<0>{})) + "x" + _TS_(g1br::at(ck_tile::number<1>{})) + "x" + _TS_(g1br::at(ck_tile::number<2>{})) + "_" +
"w" + _TS_(gwt::at(ck_tile::number<0>{})) + "x" + _TS_(gwt::at(ck_tile::number<1>{})) + "x" + _TS_(gwt::at(ck_tile::number<2>{})) + "_" +
"w" + _TS_(g0wt::at(ck_tile::number<0>{})) + "x" + _TS_(g0wt::at(ck_tile::number<1>{})) + "x" + _TS_(g0wt::at(ck_tile::number<2>{})) + "_" +
"w" + _TS_(g1wt::at(ck_tile::number<0>{})) + "x" + _TS_(g1wt::at(ck_tile::number<1>{})) + "x" + _TS_(g1wt::at(ck_tile::number<2>{})) + "_" +
(kBlockPerCuInput == -1 ? "" : ("o" + _TS_(kBlockPerCu) + "_")) + _SS_(FmhaPipeline::name) + "_" +
"v" + (std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor> ? "r" : "c") + (pn.empty() ? "" : "_" + pn) +
(BiasEnum == BlockAttentionBiasEnum::NO_BIAS ? _SS_("") : (_SS_("_") + BlockAttentionBiasEnumToStr<BiasEnum>::name)) +
(kHasMask ? "_" + _SS_(FmhaMask::name) : "") + (kDoFp8StaticQuant ? "_squant" : "") + (kIsPagedKV ? "_pagedkv" : "" );
(kHasMask ? "_" + _SS_(FmhaMask::name) : "") + (kStoreLSE ? "_lse" : "" ) + (kDoFp8StaticQuant ? "_squant" : "") + (kIsPagedKV ? "_pagedkv" : "" );
#undef _SS_
#undef _TS_
// clang-format on
......
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_combine_pipeline.hpp
View file @ 2a768425
......@@ -53,6 +53,7 @@ struct BlockFmhaFwdSplitKVCombinePipeline
using OaccDataType = remove_cvref_t<typename Problem::OaccDataType>;
using ODataType = remove_cvref_t<typename Problem::ODataType>;
static constexpr index_t kNumWarps = Problem::kNumWarps;
static constexpr index_t kBlockSize = Problem::kBlockSize;
static constexpr index_t kHeadDimV = Problem::kHeadDimV;
......@@ -117,7 +118,6 @@ struct BlockFmhaFwdSplitKVCombinePipeline
const LSEElementFunction& lse_element_func,
const OaccElementFunction& o_acc_element_func,
index_t num_splits,
index_t seqlen_q,
void* smem_ptr) const
{
// lse_acc tile in LDS
......@@ -143,11 +143,12 @@ struct BlockFmhaFwdSplitKVCombinePipeline
// copy lse_acc tile (shape=[kMaxSplits, kM0]) to LDS (shape=[kMaxSplits, kM0]).
auto lse_acc_tile = load_tile(lse_acc_dram_window);
store_tile(lse_acc_lds_write_window, lse_acc_tile);
block_sync_lds();
auto lse_accum = make_static_distributed_tensor<LSEDataType>(
Policy::template MakeLSEaccRegTileDistribution<Problem>());
__builtin_amdgcn_sched_barrier(0);
block_sync_lds();
// copy LDS (shape=[kM0, kMaxSplits]) to lse_accum (shape=[kM0, kMaxSplits])
// and fill up -INF values outside the [kM0, num_splits] region.
{
......@@ -264,46 +265,94 @@ struct BlockFmhaFwdSplitKVCombinePipeline
}
});
}
block_sync_lds();
if constexpr(kStoreLSE)
{
store_tile(lse_dram_window_tmp, tile_elementwise_in(lse_element_func, lse_logsum));
}
auto o_acc_dist = Policy::template MakeOaccDramTileDistribution<Problem>();
auto o_acc_dram_window =
auto o_acc_4_dist = Policy::template MakeOacc4DramTileDistribution<Problem>();
auto o_acc_4_dram_window =
make_tile_window(o_acc_dram_block_window_tmp.get_bottom_tensor_view(),
o_acc_dram_block_window_tmp.get_window_lengths(),
o_acc_dram_block_window_tmp.get_window_origin(),
o_acc_dist);
auto o_acc = make_static_distributed_tensor<OaccDataType>(o_acc_dist);
clear_tile(o_acc);
o_acc_4_dist);
const index_t padded_seqlen_q = integer_divide_ceil(seqlen_q, kM0) * kM0;
// shape=[4 * KM0, kN1]
auto o_acc_4 = make_static_distributed_tensor<OaccDataType>(o_acc_4_dist);
clear_tile(o_acc_4);
for(index_t i_split = 0; i_split < num_splits; ++i_split)
const index_t padded_num_splits = integer_divide_ceil(num_splits, kNumWarps) * kNumWarps;
__builtin_amdgcn_sched_barrier(0);
block_sync_lds();
// each warp handles a [KM0, kN1] tile
for(index_t split_start = 0; split_start < padded_num_splits; split_start += kNumWarps)
{
auto o_tile = load_tile(o_acc_dram_window);
auto o_tile = load_tile(o_acc_4_dram_window);
const index_t i_split = split_start + get_warp_id();
const index_t row_start = kM0 * get_warp_id();
{
constexpr auto spans = decltype(o_acc)::get_distributed_spans();
constexpr auto spans = decltype(o_acc_4)::get_distributed_spans();
sweep_tile_span(spans[number<0>{}], [&](auto idx0) {
sweep_tile_span(spans[number<1>{}], [&](auto idx1) {
constexpr auto i_j_idx = make_tuple(idx0, idx1);
const auto x_indices = get_x_indices_from_distributed_indices(
o_acc.get_tile_distribution(), i_j_idx);
o_acc_4.get_tile_distribution(), i_j_idx);
const auto row = x_indices.at(number<0>{});
const LSEDataType lse_scale = lse_acc_lds(row, i_split);
o_acc(i_j_idx) += lse_scale * o_tile(i_j_idx);
const LSEDataType lse_scale = lse_acc_lds(row - row_start, i_split);
o_acc_4(i_j_idx) += lse_scale * o_tile(i_j_idx);
});
});
}
move_tile_window(o_acc_dram_window, {padded_seqlen_q, 0});
move_tile_window(o_acc_4_dram_window, {kNumWarps * kM0, 0});
}
// 4 o_acc tiles in LDS. shape=[4 * kM0, kN1]
OaccDataType* o_acc_4_lds_ptr = static_cast<OaccDataType*>(static_cast<void*>(
static_cast<char*>(smem_ptr) + Policy::template GetSmemSizeLSEacc<Problem>()));
{
auto o_acc_4_lds_window = [&]() {
auto desc = Policy::template MakeOacc4LdsBlockDescriptor<Problem>();
auto view = make_tensor_view<address_space_enum::lds>(o_acc_4_lds_ptr, desc);
return make_tile_window(view, desc.get_lengths(), {0, 0});
}();
store_tile(o_acc_4_lds_window, o_acc_4);
}
auto o_acc_dist = Policy::template MakeOaccDramTileDistribution<Problem>();
auto o_acc_4_lds_window = [&]() {
auto desc = Policy::template MakeOacc4LdsBlockDescriptor<Problem>();
auto view = make_tensor_view<address_space_enum::lds>(o_acc_4_lds_ptr, desc);
return make_tile_window(view, desc.get_lengths(), {0, 0}, o_acc_dist);
}();
auto o_acc = make_static_distributed_tensor<OaccDataType>(o_acc_dist);
clear_tile(o_acc);
__builtin_amdgcn_sched_barrier(0);
block_sync_lds();
static_for<0, kNumWarps, 1>{}([&](auto) {
auto o_acc_in = load_tile(o_acc_4_lds_window);
{
constexpr auto spans = decltype(o_acc)::get_distributed_spans();
sweep_tile_span(spans[number<0>{}], [&](auto idx0) {
sweep_tile_span(spans[number<1>{}], [&](auto idx1) {
constexpr auto i_j_idx = make_tuple(idx0, idx1);
o_acc(i_j_idx) += o_acc_in(i_j_idx);
});
});
}
move_tile_window(o_acc_4_lds_window, {kM0, 0});
});
o_acc = tile_elementwise_in(o_acc_element_func, o_acc);
return o_acc;
......@@ -316,7 +365,6 @@ struct BlockFmhaFwdSplitKVCombinePipeline
const OaccDramBlockWindow& o_acc_dram_block_window,
LSEDramBlockWindow& lse_dram_block_window,
index_t num_splits,
index_t seqlen_q,
void* smem_ptr) const
{
return operator()(lse_acc_dram_block_window,
......@@ -325,7 +373,6 @@ struct BlockFmhaFwdSplitKVCombinePipeline
identity{},
identity{},
num_splits,
seqlen_q,
smem_ptr);
}
};
......
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_combine_pipeline_default_policy.hpp
View file @ 2a768425
......@@ -10,23 +10,38 @@ namespace ck_tile {
struct BlockFmhaFwdSplitKVCombinePipelineDefaultPolicy
{
template <index_t BlockSize, index_t M, index_t N, typename DataType>
template <index_t NumWarps, index_t M, index_t N, typename DataType>
CK_TILE_HOST_DEVICE static constexpr auto GetMaxNumWarpsForTile()
{
static_assert(NumWarps == 1 || NumWarps == 2 || NumWarps == 4);
constexpr index_t ElemPerThread = (M * N) / (NumWarps * get_warp_size());
if constexpr(0 < ElemPerThread)
{
return NumWarps;
}
else
{ // try dividing tile by smaller # of warps
return GetMaxNumWarpsForTile<NumWarps / 2, M, N, DataType>();
}
}
template <index_t NumWarps, index_t M, index_t N, typename DataType>
CK_TILE_HOST_DEVICE static constexpr auto GetVectorSizeForTile()
{
constexpr index_t PixelsPerThread = (M * N) / BlockSize;
static_assert(0 < PixelsPerThread);
constexpr index_t MaxNumWarps = GetMaxNumWarpsForTile<NumWarps, M, N, DataType>();
constexpr index_t MaxNPerThread = 16 / sizeof(DataType);
constexpr index_t NPerThread = min(MaxNPerThread, PixelsPerThread);
constexpr index_t ElemPerThread = (M * N) / (MaxNumWarps * get_warp_size());
return NPerThread;
constexpr index_t MaxNPerThread = 16 / sizeof(DataType);
return min(MaxNPerThread, ElemPerThread);
}
// alignment for dram lse tile (shape=[kMaxSplits, kM0])
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetAlignmentLSE()
{
return GetVectorSizeForTile<Problem::kBlockSize,
return GetVectorSizeForTile<Problem::kNumWarps,
Problem::kMaxSplits,
Problem::kM0,
typename Problem::LSEDataType>();
......@@ -56,40 +71,54 @@ struct BlockFmhaFwdSplitKVCombinePipelineDefaultPolicy
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSize()
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSizeLSEacc()
{
return sizeof(typename Problem::LSEDataType) *
MakeLSEaccLdsBlockDescriptor<Problem>().get_element_space_size();
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSizeOacc4()
{
return sizeof(typename Problem::OaccDataType) *
MakeOacc4LdsBlockDescriptor<Problem>().get_element_space_size();
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSize()
{
return GetSmemSizeLSEacc<Problem>() + GetSmemSizeOacc4<Problem>();
}
// shape=[kMaxSplits, kM0]
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeLSEaccDramTileDistribution()
{
using LSEDataType = remove_cvref_t<typename Problem::LSEDataType>;
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kNumWarps = Problem::kNumWarps;
constexpr index_t kNPerBlock = Problem::kM0;
constexpr index_t kMPerBlock = Problem::kMaxSplits;
constexpr index_t kNPerBlock = Problem::kM0;
constexpr index_t MaxNumWarps =
GetMaxNumWarpsForTile<Problem::kNumWarps, kNPerBlock, kMPerBlock, LSEDataType>();
constexpr index_t Replicate = Problem::kNumWarps / MaxNumWarps;
constexpr index_t NPerThread =
GetVectorSizeForTile<kBlockSize, kMPerBlock, kNPerBlock, LSEDataType>();
GetVectorSizeForTile<MaxNumWarps, kMPerBlock, kNPerBlock, LSEDataType>();
constexpr index_t NThreads = kNPerBlock / NPerThread;
constexpr index_t MThreadsPerWarp = get_warp_size() / NThreads;
constexpr index_t MPerThread = kMPerBlock / (kNumWarps * MThreadsPerWarp);
constexpr index_t MPerThread = kMPerBlock / (MaxNumWarps * MThreadsPerWarp);
static_assert(MPerThread * MaxNumWarps * MThreadsPerWarp == kMPerBlock);
static_assert(NThreads * NPerThread == kNPerBlock);
static_assert(MPerThread * kNumWarps * MThreadsPerWarp == kMPerBlock);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<MPerThread, kNumWarps, MThreadsPerWarp>,
tile_distribution_encoding<sequence<Replicate>,
tuple<sequence<MPerThread, MaxNumWarps, MThreadsPerWarp>,
sequence<NThreads, NPerThread>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<1>, sequence<2, 0>>,
tuple<sequence<0, 1>, sequence<1, 2>>,
tuple<sequence<0, 1>, sequence<2, 0>>,
sequence<1, 2>,
sequence<0, 1>>{});
}
......@@ -100,17 +129,15 @@ struct BlockFmhaFwdSplitKVCombinePipelineDefaultPolicy
{
using LSEDataType = remove_cvref_t<typename Problem::LSEDataType>;
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kMPerBlock = Problem::kMaxSplits;
constexpr index_t kNPerBlock = Problem::kM0;
constexpr index_t kMPerBlock = Problem::kM0;
constexpr index_t kNPerBlock = Problem::kMaxSplits;
constexpr index_t NPack =
GetVectorSizeForTile<kBlockSize, kMPerBlock, kNPerBlock, LSEDataType>();
GetVectorSizeForTile<Problem::kNumWarps, kMPerBlock, kNPerBlock, LSEDataType>();
constexpr auto lse_acc_lds_block_desc_0 = make_naive_tensor_descriptor(
make_tuple(number<kNPerBlock / NPack>{}, number<kMPerBlock>{}, number<NPack>{}),
make_tuple(number<(kMPerBlock + 1) * NPack>{}, number<NPack>{}, number<1>{}),
number<8>{},
number<NPack>{},
number<1>{});
constexpr auto lse_acc_lds_block_desc = transform_tensor_descriptor(
......@@ -129,17 +156,15 @@ struct BlockFmhaFwdSplitKVCombinePipelineDefaultPolicy
{
using LSEDataType = remove_cvref_t<typename Problem::LSEDataType>;
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kMPerBlock = Problem::kMaxSplits;
constexpr index_t kNPerBlock = Problem::kM0;
constexpr index_t kMPerBlock = Problem::kM0;
constexpr index_t kNPerBlock = Problem::kMaxSplits;
constexpr index_t NPack =
GetVectorSizeForTile<kBlockSize, kMPerBlock, kNPerBlock, LSEDataType>();
GetVectorSizeForTile<Problem::kNumWarps, kMPerBlock, kNPerBlock, LSEDataType>();
constexpr auto lse_acc_lds_block_desc_0 = make_naive_tensor_descriptor(
make_tuple(number<kNPerBlock / NPack>{}, number<kMPerBlock>{}, number<NPack>{}),
make_tuple(number<(kMPerBlock + 1) * NPack>{}, number<NPack>{}, number<1>{}),
number<8>{},
number<NPack>{},
number<1>{});
constexpr auto lse_acc_t_lds_block_desc = transform_tensor_descriptor(
......@@ -152,33 +177,86 @@ struct BlockFmhaFwdSplitKVCombinePipelineDefaultPolicy
return lse_acc_t_lds_block_desc;
}
// 3d + padding, shape=[4 * kM0, kN1]
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeLSEaccRegTileDistribution()
CK_TILE_HOST_DEVICE static constexpr auto MakeOacc4LdsBlockDescriptor()
{
constexpr index_t kBlockSize = Problem::kBlockSize;
using LSEDataType = remove_cvref_t<typename Problem::LSEDataType>;
constexpr index_t kNPerBlock = Problem::kMaxSplits;
constexpr index_t kMPerBlock = 4 * Problem::kM0;
constexpr index_t kNPerBlock = Problem::kN1;
constexpr index_t NPack =
GetVectorSizeForTile<Problem::kNumWarps, kMPerBlock, kNPerBlock, LSEDataType>();
constexpr auto o_acc_lds_block_desc_0 = make_naive_tensor_descriptor(
make_tuple(number<kNPerBlock / NPack>{}, number<kMPerBlock>{}, number<NPack>{}),
make_tuple(number<(kMPerBlock + 1) * NPack>{}, number<NPack>{}, number<1>{}),
number<8>{},
number<1>{});
constexpr auto o_acc_t_lds_block_desc = transform_tensor_descriptor(
o_acc_lds_block_desc_0,
make_tuple(make_pass_through_transform(kMPerBlock),
make_merge_transform(make_tuple(kNPerBlock / NPack, NPack))),
make_tuple(sequence<1>{}, sequence<0, 2>{}),
make_tuple(sequence<1>{}, sequence<0>{}));
return o_acc_t_lds_block_desc;
}
// shape=[kM0, kMaxSplits]
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeLSEaccRegTileDistribution()
{
constexpr index_t kMPerBlock = Problem::kM0;
constexpr index_t kNPerBlock = Problem::kMaxSplits;
constexpr index_t NThreads = 4;
constexpr index_t NPerThread = kNPerBlock / NThreads;
constexpr index_t MaxNThreads = 8;
constexpr index_t NThreads = min(kNPerBlock, MaxNThreads);
constexpr index_t NPerThread = kNPerBlock / NThreads;
constexpr index_t MThreads = kBlockSize / NThreads;
constexpr index_t MPerThread = kMPerBlock / MThreads;
constexpr index_t MWarps = kBlockSize / get_warp_size();
constexpr index_t MPerThread = 1;
constexpr index_t MThreads = kMPerBlock / MPerThread;
constexpr index_t MThreadPerWarp = get_warp_size() / NThreads;
constexpr index_t MaxNumWarps = (MThreads * NThreads) / get_warp_size();
constexpr index_t Replicate = Problem::kNumWarps / MaxNumWarps;
static_assert(MaxNumWarps * MThreadPerWarp * MPerThread == kMPerBlock);
static_assert(NThreads * NPerThread == kNPerBlock);
static_assert(MWarps * MThreadPerWarp * MPerThread == kMPerBlock);
return make_static_tile_distribution(
tile_distribution_encoding<
sequence<1>,
tuple<sequence<MWarps, MThreadPerWarp, MPerThread>, sequence<NThreads, NPerThread>>,
tuple<sequence<1>, sequence<2, 1>>,
tuple<sequence<0>, sequence<0, 1>>,
sequence<1, 2>,
sequence<2, 1>>{});
tile_distribution_encoding<sequence<Replicate>,
tuple<sequence<MaxNumWarps, MThreadPerWarp, MPerThread>,
sequence<NThreads, NPerThread>>,
tuple<sequence<0, 1>, sequence<2, 1>>,
tuple<sequence<0, 0>, sequence<0, 1>>,
sequence<1, 2>,
sequence<2, 1>>{});
}
// similar to MakeOaccDramTileDistribution(), but duplicate same 1-warp encoding 4 times on M
// direction
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeOacc4DramTileDistribution()
{
constexpr index_t kMPerBlock = Problem::kM0; // real kMPerBlock we want is (4 * kM0)
constexpr index_t kNPerBlock = Problem::kN1;
static_assert(get_warp_size() <= kMPerBlock * kNPerBlock);
constexpr index_t M1 = 1; // compose encoding base on 1 warp
constexpr index_t M2 = min(kMPerBlock / M1, get_warp_size());
constexpr index_t N0 = get_warp_size() / M2;
constexpr index_t N1 = kNPerBlock / N0;
constexpr index_t M0 = kMPerBlock / (M2 * M1);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<4, M0, M1, M2>, sequence<N0, N1>>,
tuple<sequence<1, 1>, sequence<1, 2>>,
tuple<sequence<0, 2>, sequence<3, 0>>,
sequence<1, 2>,
sequence<1, 1>>{});
}
template <typename Problem>
......@@ -187,6 +265,7 @@ struct BlockFmhaFwdSplitKVCombinePipelineDefaultPolicy
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kMPerBlock = Problem::kM0;
constexpr index_t kNPerBlock = Problem::kN1;
static_assert(kBlockSize <= kMPerBlock * kNPerBlock);
constexpr index_t M1 = kBlockSize / get_warp_size();
constexpr index_t M2 = min(kMPerBlock / M1, get_warp_size());
......
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_nwarp_sshuffle_qr_ks_vs.hpp 0 → 100644
View file @ 2a768425
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/fmha/block/block_attention_bias_enum.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_nwarp_sshuffle_qr_ks_vs_default_policy.hpp"
#include "ck_tile/ops/reduce/block/block_reduce.hpp"
namespace ck_tile {
// This pipeline is qkv all located in LDS
template <typename Problem_,
typename Policy_ = BlockFmhaFwdSplitKVPipelineNWarpSShuffleQRKSVSDefaultPolicy>
struct BlockFmhaFwdSplitKVPipelineNWarpSShuffleQRKSVS
{
using Problem = remove_cvref_t<Problem_>;
using Policy = remove_cvref_t<Policy_>;
using QDataType = remove_cvref_t<typename Problem::QDataType>;
using KDataType = remove_cvref_t<typename Problem::KDataType>;
using VDataType = remove_cvref_t<typename Problem::VDataType>;
using SaccDataType = remove_cvref_t<typename Problem::SaccDataType>;
using SMPLComputeDataType = remove_cvref_t<typename Problem::SMPLComputeDataType>;
using BiasDataType = remove_cvref_t<typename Problem::BiasDataType>;
using LSEDataType = remove_cvref_t<typename Problem::LSEDataType>;
using PDataType = remove_cvref_t<typename Problem::PDataType>;
using OaccDataType = remove_cvref_t<typename Problem::OaccDataType>;
using ODataType = remove_cvref_t<typename Problem::ODataType>;
using FmhaMask = remove_cvref_t<typename Problem::FmhaMask>;
using BlockFmhaShape = remove_cvref_t<typename Problem::BlockFmhaShape>;
using VLayout = remove_cvref_t<typename BlockFmhaShape::VLayout>;
static constexpr bool kQLoadOnce = true; // if q_tile load whole block length (hdim) at once
static_assert(kQLoadOnce == Policy::QLoadOnce);
static constexpr index_t kBlockSize = Problem::kBlockSize;
static constexpr index_t kM0 = BlockFmhaShape::kM0;
static constexpr index_t kN0 = BlockFmhaShape::kN0;
static constexpr index_t kK0 = BlockFmhaShape::kK0;
static constexpr index_t kN1 = BlockFmhaShape::kN1;
static constexpr index_t kK1 = BlockFmhaShape::kK1;
static constexpr index_t kQKHeaddim = BlockFmhaShape::kQKHeaddim;
static constexpr index_t kSubQKHeaddim = BlockFmhaShape::kSubQKHeaddim;
static constexpr bool kIsGroupMode = Problem::kIsGroupMode;
static constexpr bool kPadSeqLenQ = Problem::kPadSeqLenQ;
static constexpr bool kPadSeqLenK = Problem::kPadSeqLenK;
static constexpr bool kPadHeadDimQ = Problem::kPadHeadDimQ;
static constexpr bool kPadHeadDimV = Problem::kPadHeadDimV;
static constexpr auto BiasEnum = Problem::BiasEnum;
static constexpr bool kStoreLSE = Problem::kStoreLSE;
static constexpr bool kIsPagedKV = Problem::kIsPagedKV;
static constexpr bool kHasUnevenSplits = Problem::kHasUnevenSplits;
// last dimension vector length used to create tensor view(and decide buffer_load vector length)
// ... together with tensor distribution. tensor dist should able to overwrite this
static constexpr index_t kAlignmentQ =
kPadHeadDimQ ? 1 : Policy::template GetAlignmentQ<Problem>();
static constexpr index_t kAlignmentK =
kPadHeadDimQ ? 1 : Policy::template GetAlignmentK<Problem>();
static constexpr index_t kAlignmentV = []() {
if constexpr(std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor>)
return kPadHeadDimV ? 1 : Policy::template GetAlignmentV<Problem>();
else
return kPadSeqLenK ? 1 : Policy::template GetAlignmentV<Problem>();
}();
static constexpr index_t kAlignmentOacc =
kPadHeadDimV ? 1 : Policy::template GetAlignmentOacc<Problem>();
static constexpr index_t kAlignmentBias =
kPadSeqLenK ? 1 : Policy::template GetAlignmentBias<Problem>();
static constexpr index_t kBlockPerCu = []() {
if constexpr(Problem::kBlockPerCu != -1)
return Problem::kBlockPerCu;
else
{
if constexpr(kQKHeaddim <= 32)
{
return 2;
}
else if constexpr(kQKHeaddim <= 64)
{
return 3;
}
else if constexpr(kQKHeaddim <= 128)
{
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
return 1;
else
return 2;
}
else if constexpr(kQKHeaddim <= 256)
{
return 1;
}
}
}();
static constexpr const char* name = "qr_nwarp_sshuffle";
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSize()
{
return Policy::template GetSmemSize<Problem>();
}
template <typename QDramBlockWindowTmp,
typename KDramBlockWindowLengths,
typename KPageBlockNavigator,
typename VDramBlockWindowLengths,
typename VPageBlockNavigator,
typename BiasDramBlockWindowTmp,
typename LSEaccDramBlockWindowTmp,
typename QElementFunction,
typename KElementFunction,
typename VElementFunction,
typename BiasElementFunction,
typename LSEaccElementFunction,
typename SAccElementFunction,
typename PComputeElementFunction,
typename OAccElementFunction,
typename PositionEncoding>
CK_TILE_HOST_DEVICE auto
operator()(const QDramBlockWindowTmp& q_dram_block_window_tmp, // M0*K0 tile
const QElementFunction& q_element_func,
const KDramBlockWindowLengths& k_dram_block_window_lengths, // N0*K0 tile
const KPageBlockNavigator& k_page_block_navigator,
const KElementFunction& k_element_func,
const VDramBlockWindowLengths& v_dram_block_window_lengths, // N1*K1 tile
const VPageBlockNavigator& v_page_block_navigator,
const VElementFunction& v_element_func,
const BiasDramBlockWindowTmp& bias_dram_block_window_tmp, // M0*N0 tile
const BiasElementFunction& bias_element_func,
LSEaccDramBlockWindowTmp& lse_acc_dram_window_tmp, // M0*1 tile
const LSEaccElementFunction& lse_acc_element_func,
const SAccElementFunction& s_acc_element_func,
const PComputeElementFunction& p_compute_element_func,
const OAccElementFunction& o_acc_element_func,
index_t num_splits,
index_t i_split,
FmhaMask mask,
PositionEncoding position_encoding,
float scale_s,
index_t kv_l2p_offset, // logical-to-physical offset of seqlen_k coordinate
void* smem_ptr) const
{
static_assert(
std::is_same_v<QDataType, remove_cvref_t<typename QDramBlockWindowTmp::DataType>> &&
std::is_same_v<KDataType, remove_cvref_t<typename KPageBlockNavigator::DataType>> &&
std::is_same_v<VDataType, remove_cvref_t<typename VPageBlockNavigator::DataType>>,
"wrong!");
static_assert(kM0 == QDramBlockWindowTmp{}.get_window_lengths()[number<0>{}] &&
kSubQKHeaddim ==
QDramBlockWindowTmp{}.get_window_lengths()[number<1>{}] &&
kN0 == KDramBlockWindowLengths{}[number<0>{}] &&
kK0 == KDramBlockWindowLengths{}[number<1>{}] &&
kN1 == VDramBlockWindowLengths{}[number<0>{}] &&
kK1 == VDramBlockWindowLengths{}[number<1>{}] &&
kM0 == BiasDramBlockWindowTmp{}.get_window_lengths()[number<0>{}] &&
kN0 == BiasDramBlockWindowTmp{}.get_window_lengths()[number<1>{}],
"wrong!");
// Q tile in LDS
QDataType* q_lds_ptr =
static_cast<QDataType*>(static_cast<void*>(static_cast<char*>(smem_ptr)));
auto q_lds = make_tensor_view<address_space_enum::lds>(
q_lds_ptr, Policy::template MakeQLdsBlockDescriptor<Problem>());
// K tile in LDS
KDataType* k_lds_ptr =
static_cast<KDataType*>(static_cast<void*>(static_cast<char*>(smem_ptr)));
auto k_lds = make_tensor_view<address_space_enum::lds>(
k_lds_ptr, Policy::template MakeKLdsBlockDescriptor<Problem>());
auto k_lds_window =
make_tile_window(k_lds, make_tuple(number<kN0>{}, number<kK0>{}), {0, 0});
// V tile in LDS
auto v_lds = make_tensor_view<address_space_enum::lds>(
reinterpret_cast<VDataType*>(static_cast<char*>(smem_ptr) +
max(Policy::template GetSmemSizeQ<Problem>(),
Policy::template GetSmemSizeK<Problem>())),
Policy::template MakeVLdsBlockDescriptor<Problem>());
auto v_lds_window = make_tile_window(
v_lds, Policy::template MakeVLdsBlockDescriptor<Problem>().get_lengths(), {0, 0});
// S tile in LDS
auto s_lds = make_tensor_view<address_space_enum::lds>(
reinterpret_cast<SaccDataType*>(reinterpret_cast<char*>(smem_ptr) +
max(Policy::template GetSmemSizeQ<Problem>(),
Policy::template GetSmemSizeK<Problem>())),
Policy::template MakeSLdsBlockDescriptor<Problem>());
auto s_write_lds_window = make_tile_window(
s_lds, Policy::template MakeSLdsBlockDescriptor<Problem>().get_lengths(), {0, 0});
auto s_read_lds_window =
make_tile_window(s_lds,
Policy::template MakeSLdsBlockDescriptor<Problem>().get_lengths(),
{0, 0},
Policy::template MakeSRegTileDistribution<Problem>());
// Block GEMM
constexpr auto gemm_0 = Policy::template GetQKBlockGemm<Problem>();
constexpr auto gemm_1 = Policy::template GetKVBlockGemm<Problem>();
auto q_dram_window =
make_tile_window(q_dram_block_window_tmp.get_bottom_tensor_view(),
q_dram_block_window_tmp.get_window_lengths(),
q_dram_block_window_tmp.get_window_origin(),
Policy::template MakeQDramTileDistribution<Problem>());
// load Q here, will store Q into LDS to maximize throughput
auto origin_q = load_tile(q_dram_window);
using SaccBlockTileType = decltype(gemm_0.MakeCBlockTile());
auto s_acc = SaccBlockTileType{};
// reduction function for softmax
const auto f_max = [](auto e0, auto e1) { return max(e0, e1); };
const auto f_sum = [](auto e0, auto e1) { return e0 + e1; };
using OaccBlockTileType = decltype(gemm_1.MakeCBlockTile());
auto o_acc = OaccBlockTileType{};
// infer Sacc, S, P, M, L, Oacc type
using SBlockTileType = decltype(cast_tile<SMPLComputeDataType>(o_acc));
using MLBlockTileType = decltype(block_tile_reduce<SMPLComputeDataType>(
SBlockTileType{}, sequence<1>{}, f_max, SMPLComputeDataType{0}));
// init M, L
auto m = MLBlockTileType{};
auto l = MLBlockTileType{};
clear_tile(o_acc);
set_tile(m, -numeric<SMPLComputeDataType>::infinity());
clear_tile(l);
const auto q_origin = q_dram_window.get_window_origin();
const auto [logical_seqlen_k_start, logical_seqlen_k_end] = mask.GetTileRangeAlongX(
q_origin.at(number<0>{}), number<kM0>{}, number<kN0>{}, num_splits, i_split);
// check early exit if no work to do
if constexpr(FmhaMask::IsMasking || kPadSeqLenK || kHasUnevenSplits)
{
const index_t logical_num_total_loop =
integer_divide_ceil(logical_seqlen_k_end - logical_seqlen_k_start, kN0);
if(logical_num_total_loop <= 0)
{
if constexpr(kStoreLSE)
{
auto lse_acc =
make_static_distributed_tensor<LSEDataType>(m.get_tile_distribution());
set_tile(lse_acc, -numeric<SMPLComputeDataType>::infinity());
if(get_thread_local_1d_id() < kM0)
{
store_tile(lse_acc_dram_window_tmp,
tile_elementwise_in(lse_acc_element_func, lse_acc));
}
}
// Note: here occ are all cleard, return it
// Note: q loaded but no fence, ignore it.
return o_acc;
}
}
const index_t physical_seqlen_k_start = logical_seqlen_k_start + kv_l2p_offset;
const index_t physical_seqlen_k_end = logical_seqlen_k_end + kv_l2p_offset;
// make sure the first tile is completely located in page-block (page-block size should be
// divisible by kN0)
// relationship between each *_start variables: aligned_physical_seqlen_k_start <=
// physical_seqlen_k_start, logical_seqlen_k_start <= physical_seqlen_k_start
const index_t aligned_physical_seqlen_k_start =
[&, physical_seqlen_k_start_ = physical_seqlen_k_start] {
if constexpr(kIsPagedKV)
{
return kN0 * integer_divide_floor(physical_seqlen_k_start_, kN0);
}
else
{
return physical_seqlen_k_start_;
}
}();
const index_t num_total_loop =
integer_divide_ceil(physical_seqlen_k_end - aligned_physical_seqlen_k_start, kN0);
auto [i_page_block_k, k_dram_block_window] = k_page_block_navigator.make_tile_window(
k_dram_block_window_lengths, {aligned_physical_seqlen_k_start, 0});
const auto bias_origin = bias_dram_block_window_tmp.get_window_origin();
auto bias_dram_window =
make_tile_window(bias_dram_block_window_tmp.get_bottom_tensor_view(),
bias_dram_block_window_tmp.get_window_lengths(),
{bias_origin.at(number<0>{}),
logical_seqlen_k_start - (physical_seqlen_k_start -
aligned_physical_seqlen_k_start)}, // M/N
Policy::template MakeBiasDramTileDistribution<decltype(gemm_0)>());
auto [i_page_block_v, v_dram_window] = v_page_block_navigator.make_tile_window(
v_dram_block_window_lengths,
{0, aligned_physical_seqlen_k_start}, // TODO: hdim split?
Policy::template MakeVDramTileDistribution<Problem>());
// store Q into LDS
__builtin_amdgcn_sched_barrier(0);
auto q_lds_window_for_store = make_tile_window(
q_lds, Policy::template MakeQLdsBlockDescriptor<Problem>().get_lengths(), {0, 0});
store_tile(q_lds_window_for_store, origin_q);
__builtin_amdgcn_sched_barrier(0);
// load Q from LDS
__builtin_amdgcn_sched_barrier(0);
auto q_lds_window_for_load = make_tile_window(
q_lds,
Policy::template MakeQLdsBlockDescriptor<Problem>().get_lengths(),
{0, 0},
Policy::template MakeQRegTileDistribution<Problem, decltype(gemm_0)>());
block_sync_lds();
auto q = load_tile(q_lds_window_for_load);
__builtin_amdgcn_sched_barrier(0);
auto q_tile = tile_elementwise_in(q_element_func, q);
// prefetch K tile
index_t i_total_loops = 0;
constexpr index_t k0_loops = kQKHeaddim / kK0;
constexpr index_t k1_loops = kN0 / kK1;
static_assert(2 <= k0_loops);
static_assert(1 <= k1_loops);
auto k_dram_window = make_tile_window(
k_dram_block_window,
Policy::template MakeKDramTileDistribution<Problem>()); // K DRAM tile window for
// load the first tile of the first iteration and store to LDS
auto k_block_tile = load_tile(k_dram_window);
// moving k_dram_window is an in-page-block operation, so there is
// no need to invoke k_page_block_navigator.move_tile_window() here.
move_tile_window(k_dram_window, {0, kK0});
store_tile(k_lds_window, tile_elementwise_in(k_element_func, k_block_tile));
do
{
// STAGE 1, QK gemm
clear_tile(s_acc); // initialize C
// load the second tile of the first iteration
k_block_tile = load_tile(k_dram_window);
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
{
__builtin_amdgcn_sched_barrier(
0); // prevent from messing up the order of global loads
}
const auto bias_tile = load_tile(bias_dram_window); // load bias tile
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
{
__builtin_amdgcn_sched_barrier(
0); // prevent from messing up the order of global loads
}
if constexpr(k0_loops > 2)
{
static_for<0, k0_loops - 2, 1>{}([&](auto i_k0) {
block_sync_lds();
gemm_0(s_acc,
get_slice_tile(q_tile,
sequence<0, i_k0 * kK0>{},
sequence<kM0, (i_k0 + 1) * kK0>{}),
k_lds_window);
block_sync_lds();
move_tile_window(k_dram_window, {0, kK0});
store_tile(
k_lds_window,
tile_elementwise_in(k_element_func, k_block_tile)); // LDS write i + 1
k_block_tile = load_tile(k_dram_window); // global read i + 2
});
}
const auto v_prefetch = load_tile(v_dram_window); // prefetch load v tile
{ // tail
block_sync_lds();
gemm_0(s_acc,
get_slice_tile(q_tile,
sequence<0, (k0_loops - 2) * kK0>{},
sequence<kM0, (k0_loops - 1) * kK0>{}),
k_lds_window);
block_sync_lds();
store_tile(k_lds_window, tile_elementwise_in(k_element_func, k_block_tile));
block_sync_lds();
gemm_0(s_acc,
get_slice_tile(q_tile,
sequence<0, (k0_loops - 1) * kK0>{},
sequence<kM0, k0_loops * kK0>{}),
k_lds_window);
}
// STAGE 2, scale_s, add bias, mask, softmax
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
{
s_acc = tile_elementwise_in(s_acc_element_func, s_acc);
tile_elementwise_inout([&scale_s](auto& x) { x = x * scale_s; }, s_acc);
tile_elementwise_inout(
[&](auto& x, const auto& y) {
#if !CK_TILE_FMHA_FWD_FAST_EXP2
x += type_convert<SaccDataType>(bias_element_func(y));
#else
x += log2e_v<SaccDataType> *
type_convert<SaccDataType>(bias_element_func(y));
#endif
},
s_acc,
bias_tile);
}
else if constexpr(BiasEnum == BlockAttentionBiasEnum::ALIBI)
{
const auto k_origin = k_page_block_navigator.to_global_window_origin(
i_page_block_k, k_dram_block_window.get_window_origin());
constexpr auto s_spans = decltype(s_acc)::get_distributed_spans();
s_acc = tile_elementwise_in(s_acc_element_func, s_acc);
sweep_tile_span(s_spans[number<0>{}], [&](auto idx0) {
sweep_tile_span(s_spans[number<1>{}], [&](auto idx1) {
const auto tile_idx = get_x_indices_from_distributed_indices(
s_acc.get_tile_distribution(), make_tuple(idx0, idx1));
const auto row = q_origin.at(number<0>{}) + tile_idx.at(number<0>{});
const auto col = k_origin.at(number<0>{}) + tile_idx.at(number<1>{});
constexpr auto i_j_idx = make_tuple(idx0, idx1);
s_acc(i_j_idx) *= scale_s;
// position_encoding accept only logical coordinates, do conversion here
position_encoding.update(s_acc(i_j_idx), row, col - kv_l2p_offset);
});
});
}
else
{
s_acc = tile_elementwise_in(s_acc_element_func, s_acc);
#if !CK_TILE_FMHA_FWD_FAST_EXP2
tile_elementwise_inout([&scale_s](auto& x) { x = x * scale_s; }, s_acc);
#endif
}
move_tile_window(bias_dram_window, {0, kN0});
/// TODO: only check in first/last iteration without increasing code size
if constexpr(kHasUnevenSplits)
{
const auto k_origin = k_page_block_navigator.to_global_window_origin(
i_page_block_k, k_dram_block_window.get_window_origin());
set_tile_if(
s_acc,
-numeric<SMPLComputeDataType>::infinity(),
[&,
physical_seqlen_k_start_ = physical_seqlen_k_start,
physical_seqlen_k_end_ = physical_seqlen_k_end](auto tile_idx) {
const auto col = k_origin.at(number<0>{}) + tile_idx.at(number<1>{});
if constexpr(kIsPagedKV)
{
return col < physical_seqlen_k_start_ || physical_seqlen_k_end_ <= col;
}
else
{
return physical_seqlen_k_end_ <= col;
}
});
}
if constexpr(kPadSeqLenK || FmhaMask::IsMasking)
{
const auto k_origin = k_page_block_navigator.to_global_window_origin(
i_page_block_k, k_dram_block_window.get_window_origin());
// mask accept only logical coordinates, do conversion here
bool need_perpixel_check = mask.IsEdgeTile(q_origin.at(number<0>{}),
k_origin.at(number<0>{}) - kv_l2p_offset,
number<kM0>{},
number<kN0>{});
if(need_perpixel_check)
{
set_tile_if(
s_acc, -numeric<SMPLComputeDataType>::infinity(), [&](auto tile_idx) {
const auto row = q_origin.at(number<0>{}) + tile_idx.at(number<0>{});
const auto col = k_origin.at(number<0>{}) + tile_idx.at(number<1>{});
return mask.IsOutOfBound(row, col - kv_l2p_offset);
});
}
}
__builtin_amdgcn_sched_barrier(0);
// load the first tile for next iteration
if(i_total_loops < num_total_loop - 1)
{
// move K tile windows
i_page_block_k = k_page_block_navigator.move_tile_window(
i_page_block_k, k_dram_block_window, {kN0, 0});
k_dram_window = make_tile_window(
k_dram_block_window,
Policy::template MakeKDramTileDistribution<Problem>()); // K DRAM tile window
// laod the first tile of the first iteration and store to LDS
k_block_tile = load_tile(k_dram_window);
}
__builtin_amdgcn_sched_barrier(0);
const auto s = cast_tile<SMPLComputeDataType>(s_acc); // S{j}
// shuffle through LDS so that the tile layout is consistent with required by Gemm1
store_tile(s_write_lds_window, s);
block_sync_lds();
auto s_new = load_tile(s_read_lds_window);
auto m_local = block_tile_reduce<SMPLComputeDataType>(
s_new,
sequence<1>{},
f_max,
-numeric<SMPLComputeDataType>::infinity()); // m_local = rowmax(S{j})
block_tile_reduce_sync(m_local, f_max, bool_constant<false>{});
const auto m_old = m; // m{j-1}
tile_elementwise_inout(
[](auto& e0, auto e1, auto e2) { e0 = max(e1, e2); }, m, m_old, m_local); // m{j}
auto p_compute = make_static_distributed_tensor<SMPLComputeDataType>(
s_new.get_tile_distribution()); // Pcompute{j}
static const auto get_validated_m = [](SMPLComputeDataType raw_m) {
/// NOTICE: bias might be materialized mask including -inf values, need
/// consideration
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS ||
FmhaMask::IsMasking)
{
return raw_m == -numeric<SMPLComputeDataType>::infinity()
? type_convert<SMPLComputeDataType>(0.f)
: raw_m;
}
else
{
return raw_m;
}
};
constexpr auto p_spans = decltype(p_compute)::get_distributed_spans();
sweep_tile_span(p_spans[number<0>{}], [&](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
#if CK_TILE_FMHA_FWD_FAST_EXP2
auto row_max = scale_s * get_validated_m(m[i_idx]);
#endif
sweep_tile_span(p_spans[number<1>{}], [&](auto idx1) {
constexpr auto i_j_idx = make_tuple(idx0, idx1);
#if CK_TILE_FMHA_FWD_FAST_EXP2
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS ||
BiasEnum == BlockAttentionBiasEnum::ALIBI)
{
p_compute(i_j_idx) = exp2(s_new[i_j_idx] - get_validated_m(m[i_idx]));
}
else
{
p_compute(i_j_idx) = exp2(scale_s * s_new[i_j_idx] - row_max);
}
#else
p_compute(i_j_idx) = exp(s_new[i_j_idx] - get_validated_m(m[i_idx]));
#endif
});
});
auto rowsum_p = block_tile_reduce<SMPLComputeDataType>(
p_compute, sequence<1>{}, f_sum, SMPLComputeDataType{0}); // rowsum(Pcompute{j})
block_tile_reduce_sync(rowsum_p, f_sum, bool_constant<false>{});
const auto p =
cast_tile<PDataType>(tile_elementwise_in(p_compute_element_func, p_compute));
// l{j}, Oacc{j}
constexpr auto o_spans = decltype(o_acc)::get_distributed_spans();
sweep_tile_span(o_spans[number<0>{}], [&](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
#if CK_TILE_FMHA_FWD_FAST_EXP2
const auto tmp = [&]() {
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS ||
BiasEnum == BlockAttentionBiasEnum::ALIBI)
{
return exp2(m_old[i_idx] - get_validated_m(m[i_idx]));
}
else
{
auto row_max = scale_s * get_validated_m(m[i_idx]);
return exp2(scale_s * m_old[i_idx] - row_max);
}
}();
#else
const auto tmp = exp(m_old[i_idx] - get_validated_m(m[i_idx]));
#endif
l(i_idx) = tmp * l[i_idx] + rowsum_p[i_idx];
sweep_tile_span(o_spans[number<1>{}], [&](auto idx1) {
constexpr auto i_j_idx = make_tuple(idx0, idx1);
// FIXME: this use different equation from FA v2 paper,
// but produce correc result.
// Is the equation wrong?
o_acc(i_j_idx) *= tmp;
});
});
block_sync_lds();
if constexpr(std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor>)
{
auto v_shuffle_tmp = make_static_distributed_tensor<VDataType>(
Policy::template MakeShuffledVRegBlockDescriptor<Problem>());
shuffle_tile(v_shuffle_tmp, v_prefetch);
store_tile(
v_lds_window,
tile_elementwise_in(v_element_func, v_shuffle_tmp)); // store the prefetch
}
else
{
store_tile(v_lds_window,
tile_elementwise_in(v_element_func, v_prefetch)); // store the prefetch
}
i_page_block_v =
v_page_block_navigator.move_tile_window(i_page_block_v, v_dram_window, {0, kK1});
// STAGE 3, KV gemm
if constexpr(k1_loops > 1)
{
static_for<0, k1_loops - 1, 1>{}([&,
&i_page_block_v_ = i_page_block_v,
&v_dram_window_ = v_dram_window](auto i_k1) {
const auto v = load_tile(v_dram_window_); // load next v
block_sync_lds();
gemm_1(o_acc,
get_slice_tile(
p, sequence<0, i_k1 * kK1>{}, sequence<kM0, (i_k1 + 1) * kK1>{}),
v_lds_window);
block_sync_lds();
if constexpr(std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor>)
{
auto v_shuffle_tmp = make_static_distributed_tensor<VDataType>(
Policy::template MakeShuffledVRegBlockDescriptor<Problem>());
shuffle_tile(v_shuffle_tmp, v);
store_tile(v_lds_window,
tile_elementwise_in(v_element_func,
v_shuffle_tmp)); // store the prefetch
}
else
{
store_tile(v_lds_window,
tile_elementwise_in(v_element_func, v)); // store next v
}
i_page_block_v_ = v_page_block_navigator.move_tile_window(
i_page_block_v_, v_dram_window_, {0, kK1});
});
}
// tail
{
block_sync_lds();
gemm_1(o_acc,
get_slice_tile(
p, sequence<0, (k1_loops - 1) * kK1>{}, sequence<kM0, k1_loops * kK1>{}),
v_lds_window);
block_sync_lds();
}
__builtin_amdgcn_sched_barrier(0);
// load the first tile for next iteration
if(i_total_loops < num_total_loop - 1)
{
// store the first tile for next iteration to LDS
// moving k_dram_window is an in-page-block operation, so there is
// no need to invoke k_page_block_navigator.move_tile_window() here.
move_tile_window(k_dram_window, {0, kK0});
store_tile(k_lds_window, tile_elementwise_in(k_element_func, k_block_tile));
}
} while(++i_total_loops < num_total_loop);
if constexpr(kStoreLSE)
{
// store lse acc
auto lse_acc = make_static_distributed_tensor<LSEDataType>(m.get_tile_distribution());
constexpr auto lse_acc_spans = decltype(lse_acc)::get_distributed_spans();
sweep_tile_span(lse_acc_spans[number<0>{}], [&, m_ = m, l_ = l](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
#if CK_TILE_FMHA_FWD_FAST_EXP2
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS ||
BiasEnum == BlockAttentionBiasEnum::ALIBI)
{
lse_acc(i_idx) = m_[i_idx] / C_LOG2E + log(l_[i_idx]);
}
else
{
lse_acc(i_idx) = m_[i_idx] * scale_s / C_LOG2E + log(l_[i_idx]);
}
#else
lse_acc(i_idx) = m_[i_idx] + log(l_[i_idx]);
#endif
});
if(get_thread_local_1d_id() < kM0)
{
store_tile(lse_acc_dram_window_tmp,
tile_elementwise_in(lse_acc_element_func, lse_acc));
}
}
// finally, O
constexpr auto o_spans = decltype(o_acc)::get_distributed_spans();
sweep_tile_span(o_spans[number<0>{}], [&](auto idx0) {
constexpr auto i_idx = make_tuple(idx0);
const auto tmp = [&]() {
if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS ||
FmhaMask::IsMasking)
{
return l[i_idx] == 0.f ? 0.f : 1 / l[i_idx];
}
else
return 1 / l[i_idx];
}();
sweep_tile_span(o_spans[number<1>{}], [&](auto idx1) {
constexpr auto i_j_idx = make_tuple(idx0, idx1);
o_acc(i_j_idx) *= tmp;
});
});
o_acc = tile_elementwise_in(o_acc_element_func, o_acc);
return o_acc;
}
template <typename QDramBlockWindowTmp,
typename KDramBlockWindowLengths,
typename KPageBlockNavigator,
typename VDramBlockWindowLengths,
typename VPageBlockNavigator,
typename BiasDramBlockWindowTmp,
typename LSEaccDramBlockWindowTmp,
typename PositionEncoding>
CK_TILE_HOST_DEVICE auto
operator()(const QDramBlockWindowTmp& q_dram_block_window_tmp, // M0*K0 tile
const KDramBlockWindowLengths& k_dram_block_window_lengths, // N0*K0 tile
const KPageBlockNavigator& k_page_block_navigator,
const VDramBlockWindowLengths& v_dram_block_window_lengths, // N1*K1 tile
const VPageBlockNavigator& v_page_block_navigator,
const BiasDramBlockWindowTmp& bias_dram_block_window_tmp, // M0*N0 tile
LSEaccDramBlockWindowTmp& lse_acc_dram_block_window_tmp, // M0*1 tile
index_t num_splits,
index_t i_split,
FmhaMask mask,
PositionEncoding position_encoding,
float scale_s,
index_t kv_l2p_offset, // logical-to-physical offset of seqlen_k coordinate
void* smem_ptr) const
{
return operator()(q_dram_block_window_tmp,
identity{},
k_dram_block_window_lengths,
k_page_block_navigator,
identity{},
v_dram_block_window_lengths,
v_page_block_navigator,
identity{},
bias_dram_block_window_tmp,
identity{},
lse_acc_dram_block_window_tmp,
identity{},
identity{},
identity{},
identity{},
num_splits,
i_split,
mask,
position_encoding,
scale_s,
kv_l2p_offset,
smem_ptr);
}
};
} // namespace ck_tile
include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_nwarp_sshuffle_qr_ks_vs_default_policy.hpp 0 → 100644
View file @ 2a768425
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qx_ks_vs_custom_policy.hpp"
#include "ck_tile/ops/gemm/block/block_gemm_asmem_bsmem_creg_v1_custom_policy.hpp"
#include "ck_tile/ops/gemm/block/block_gemm_asmem_bsmem_creg_v1.hpp"
namespace ck_tile {
// This pipeline is qkv all located in LDS
struct BlockFmhaFwdSplitKVPipelineNWarpSShuffleQRKSVSDefaultPolicy
: BlockFmhaPipelineQXKSVSCustomPolicy</* QLoadOnce = */ true,
/* AsyncCopyK = */ false,
/* AsyncCopyV = */ false,
/* NumPrefetchK = */ 1,
/* NumPrefetchV = */ 1>
{
using BasePolicy = BlockFmhaPipelineQXKSVSCustomPolicy</* QLoadOnce = */ true,
/* AsyncCopyK = */ false,
/* AsyncCopyV = */ false,
/* NumPrefetchK = */ 1,
/* NumPrefetchV = */ 1>;
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetAlignmentQ()
{
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kMPerBlock = Problem::BlockFmhaShape::kM0;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kSubQKHeaddim;
constexpr index_t MaxVectorSize = 16 / sizeof(typename Problem::QDataType);
// this should align with MakeQDramTileDistribution()
constexpr index_t ElemPerThread = (kMPerBlock * kKPerBlock) / kBlockSize;
static_assert(0 < ElemPerThread);
return min(ElemPerThread, MaxVectorSize);
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetAlignmentOacc()
{
using OaccDataType = remove_cvref_t<typename Problem::OaccDataType>;
return static_cast<index_t>(16 / sizeof(OaccDataType));
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeQDramTileDistribution()
{
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kMPerBlock = Problem::BlockFmhaShape::kM0;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kSubQKHeaddim;
constexpr index_t MaxVectorSize = 16 / sizeof(typename Problem::QDataType);
constexpr index_t ElemPerThread = (kMPerBlock * kKPerBlock) / kBlockSize;
static_assert(0 < ElemPerThread);
constexpr index_t kMaxVecLoad = min(ElemPerThread, MaxVectorSize);
constexpr index_t KPerThread = kMaxVecLoad;
constexpr index_t KThreads = kKPerBlock / KPerThread;
constexpr index_t MThreadPerWarp = get_warp_size() / KThreads;
constexpr index_t NumWarps = kBlockSize / get_warp_size();
constexpr index_t MPerThread = kMPerBlock / (MThreadPerWarp * NumWarps);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<MPerThread, NumWarps, MThreadPerWarp>,
sequence<KThreads, KPerThread>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<1>, sequence<2, 0>>,
sequence<1, 2>,
sequence<0, 1>>{});
}
template <typename Problem, typename BlockGemm>
CK_TILE_HOST_DEVICE static constexpr auto MakeQRegTileDistribution()
{
return BasePolicy::template MakeQDramTileDistribution<Problem, BlockGemm>();
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetSmemKPackQ()
{
// TODO: this is for 3d layout
using QDataType = remove_cvref_t<typename Problem::QDataType>;
return static_cast<index_t>(16 / sizeof(QDataType));
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeQLdsBlockDescriptor()
{
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kMPerBlock = Problem::BlockFmhaShape::kM0;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kSubQKHeaddim;
constexpr index_t ElemPerThread = (kMPerBlock * kKPerBlock) / kBlockSize;
static_assert(0 < ElemPerThread);
constexpr index_t kKPack = min(ElemPerThread, GetSmemKPackQ<Problem>());
constexpr auto q_lds_block_desc_0 = make_naive_tensor_descriptor(
make_tuple(number<kKPerBlock / kKPack>{}, number<kMPerBlock>{}, number<kKPack>{}),
make_tuple(number<(kMPerBlock + 1) * kKPack>{}, number<kKPack>{}, number<1>{}),
number<kKPack>{},
number<1>{});
constexpr auto q_lds_block_desc = transform_tensor_descriptor(
q_lds_block_desc_0,
make_tuple(
make_pass_through_transform(number<kMPerBlock>{}),
make_merge_transform(make_tuple(number<kKPerBlock / kKPack>{}, number<kKPack>{}))),
make_tuple(sequence<1>{}, sequence<0, 2>{}),
make_tuple(sequence<0>{}, sequence<1>{}));
return q_lds_block_desc;
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetSmemNPackS()
{
using SDataType = remove_cvref_t<typename Problem::SaccDataType>;
return static_cast<index_t>(16 / sizeof(SDataType));
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeSLdsBlockDescriptor()
{
constexpr index_t kMPerBlock = Problem::BlockFmhaShape::kM0;
constexpr index_t kNPerBlock = Problem::BlockFmhaShape::kN0;
constexpr index_t kNPack = GetSmemNPackS<Problem>();
constexpr auto s_lds_block_desc_0 = make_naive_tensor_descriptor(
make_tuple(number<kNPerBlock / kNPack>{}, number<kMPerBlock>{}, number<kNPack>{}),
make_tuple(number<(kMPerBlock + 1) * kNPack>{}, number<kNPack>{}, number<1>{}),
number<kNPack>{},
number<1>{});
constexpr auto s_lds_block_desc = transform_tensor_descriptor(
s_lds_block_desc_0,
make_tuple(
make_pass_through_transform(number<kMPerBlock>{}),
make_merge_transform(make_tuple(number<kNPerBlock / kNPack>{}, number<kNPack>{}))),
make_tuple(sequence<1>{}, sequence<0, 2>{}),
make_tuple(sequence<0>{}, sequence<1>{}));
return s_lds_block_desc;
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeSRegTileDistribution()
{
using BlockGemm = remove_cvref_t<decltype(GetKVBlockGemm<Problem>())>;
constexpr auto config = BlockGemm::Policy::template GetWarpGemmMWarpNWarp<Problem>();
using WG = remove_cvref_t<decltype(config.template at<0>())>;
constexpr index_t MWarp = config.template at<1>();
constexpr index_t NWarp = config.template at<2>();
static_assert(MWarp == 1, "Check failed!");
constexpr index_t kMPerBlock = Problem::BlockFmhaShape::kM0;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kK1;
constexpr index_t kTileK = Problem::BlockFmhaShape::kN0;
// K2 is equal to Impl::kABKPerLane * kKIterPerWarpGemm
constexpr index_t K3 = WG::kK / WG::WarpGemmAttribute::Impl::kABKLane;
constexpr index_t K2 = WG::WarpGemmAttribute::Impl::kABKLane;
constexpr index_t K1 = kKPerBlock / (K2 * K3);
constexpr index_t K0 = kTileK / kKPerBlock;
constexpr index_t M2 = WG::WarpGemmAttribute::Impl::kAMLane;
constexpr index_t M1 = MWarp;
constexpr index_t M0 = kMPerBlock / (M2 * M1);
constexpr auto s2_block_dstr_encoding =
tile_distribution_encoding<sequence<NWarp>,
tuple<sequence<M0, M1, M2>, sequence<K0, K1, K2, K3>>,
tuple<sequence<1, 0>, sequence<2, 1>>,
tuple<sequence<1, 0>, sequence<2, 2>>,
sequence<1, 2, 2, 2>,
sequence<0, 0, 1, 3>>{};
constexpr auto s2_block_dstr = make_static_tile_distribution(s2_block_dstr_encoding);
return s2_block_dstr;
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSizeQ()
{
return MakeQLdsBlockDescriptor<Problem>().get_element_space_size() *
sizeof(typename Problem::QDataType);
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSizeK()
{
return MakeKLdsBlockDescriptor<Problem>().get_element_space_size() *
sizeof(typename Problem::KDataType);
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSizeV()
{
return MakeVLdsBlockDescriptor<Problem>().get_element_space_size() *
sizeof(typename Problem::VDataType);
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSizeS()
{
return MakeSLdsBlockDescriptor<Problem>().get_element_space_size() *
sizeof(typename Problem::SaccDataType);
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSize()
{
return max(GetSmemSizeQ<Problem>(), GetSmemSizeK<Problem>()) +
max(GetSmemSizeV<Problem>(), GetSmemSizeS<Problem>());
}
};
} // namespace ck_tile
include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_problem.hpp
View file @ 2a768425
......@@ -106,28 +106,43 @@ struct BlockFmhaFwdSplitKVPipelineProblem
static constexpr index_t kBlockPerCu = Traits::kBlockPerCu;
};
// extract tile size attributes to remove dependency on traits
template <typename OaccDataType_, ck_tile::index_t kN1_>
struct BlockFmhaSplitKVCombinePipelineTileSizes
{
static constexpr index_t MaxVectorSize = 16 / sizeof(OaccDataType_);
static constexpr index_t kN1 = kN1_;
static constexpr index_t NThreads = kN1 / MaxVectorSize;
static constexpr index_t kM0 = get_warp_size() / NThreads; // MThreadPerWarp
};
template <typename LSEDataType_,
typename OaccDataType_,
typename ODataType_,
index_t HeadDimV_,
index_t kM0_,
index_t kN1_,
bool kIsGroupMode_,
ck_tile::index_t kN1_,
typename Traits_>
struct BlockFmhaSplitKVCombinePipelineProblem
: BlockFmhaSplitKVCombinePipelineTileSizes<OaccDataType_, kN1_>
{
using BaseType = BlockFmhaSplitKVCombinePipelineTileSizes<OaccDataType_, kN1_>;
using LSEDataType = remove_cvref_t<LSEDataType_>;
using OaccDataType = remove_cvref_t<OaccDataType_>;
using ODataType = remove_cvref_t<ODataType_>;
using Traits = remove_cvref_t<Traits_>;
static constexpr index_t kNumWarps = kM0_ / (get_warp_size() / 4);
static constexpr index_t kBlockSize = kNumWarps * get_warp_size();
static constexpr bool kIsGroupMode = kIsGroupMode_;
static_assert(std::is_same_v<LSEDataType, OaccDataType>);
static constexpr index_t kHeadDimV = HeadDimV_;
static constexpr index_t kM0 = kM0_;
static constexpr index_t kN1 = kN1_;
static constexpr bool kIsGroupMode = kIsGroupMode_;
using BaseType::kM0;
using BaseType::kN1;
static_assert(kN1 <= kHeadDimV && kHeadDimV % kN1 == 0);
// attributes from traits
static constexpr bool kPadSeqLenQ = Traits::kPadSeqLenQ;
......@@ -136,6 +151,13 @@ struct BlockFmhaSplitKVCombinePipelineProblem
static constexpr bool kDoFp8StaticQuant = Traits::kDoFp8StaticQuant;
static constexpr index_t kBlockPerCu = Traits::kBlockPerCu;
static constexpr index_t kMaxSplits = Traits::kMaxSplits;
static_assert(8 <= kMaxSplits);
static constexpr index_t kNumWarps = 4; // always use 4 warps for each workgroup
static constexpr index_t kBlockSize = kNumWarps * get_warp_size();
static_assert(get_warp_size() <= (kM0 * kMaxSplits) &&
(kM0 * kMaxSplits) % get_warp_size() == 0);
};
template <typename QDataType_,
......
include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qx_ks_vs_custom_policy.hpp
View file @ 2a768425
......@@ -41,52 +41,21 @@ struct BlockFmhaPipelineQXCustomPolicy</* QLoadOnce = */ true>
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetAlignmentQ()
{
constexpr index_t MaxVectorSize = 16 / sizeof(typename Problem::QDataType);
using BlockGemm = remove_cvref_t<decltype(GetQKBlockGemm<Problem>())>;
constexpr auto config = BlockGemm::Policy::template GetWarpGemmMWarpNWarp<Problem>();
using WG = remove_cvref_t<decltype(config.template at<0>())>;
return WG::kK / WG::WarpGemmAttribute::Impl::kABKLane;
return min(MaxVectorSize, WG::kK / WG::WarpGemmAttribute::Impl::kABKLane);
}
template <typename Problem, typename BlockGemm>
CK_TILE_HOST_DEVICE static constexpr auto MakeQDramTileDistribution()
{
constexpr auto config = BlockGemm::Policy::template GetWarpGemmMWarpNWarp<Problem>();
using WG = remove_cvref_t<decltype(config.template at<0>())>;
constexpr index_t MWarp = config.template at<1>();
constexpr index_t kMPerBlock = Problem::BlockFmhaShape::kM0;
constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kSubQKHeaddim;
constexpr index_t K2 = WG::kK / WG::WarpGemmAttribute::Impl::kABKLane;
constexpr index_t K1 = WG::WarpGemmAttribute::Impl::kABKLane;
constexpr index_t K0 = kKPerBlock / (K1 * K2);
constexpr index_t M2 = WG::WarpGemmAttribute::Impl::kAMLane;
constexpr index_t M1 = MWarp;
constexpr index_t M0 = kMPerBlock / (M2 * M1);
if constexpr(1 < Problem::kNumGemm0Warps)
{
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1, M2>, sequence<K0, K1, K2>>,
tuple<sequence<1>, sequence<2, 1>>,
tuple<sequence<1>, sequence<1, 2>>,
sequence<1, 2, 2>,
sequence<0, 0, 2>>{});
}
else
{
static_assert(MWarp == 1);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1, M2>, sequence<K0, K1, K2>>,
tuple<sequence<2, 1>>,
tuple<sequence<1, 2>>,
sequence<1, 2, 2>,
sequence<0, 0, 2>>{});
}
return BlockGemm::template MakeABlockTileDistribution<
Problem::BlockFmhaShape::kM0,
Problem::BlockFmhaShape::kSubQKHeaddim>();
}
template <typename Problem>
......@@ -105,7 +74,7 @@ struct BlockFmhaPipelineQXCustomPolicy</* QLoadOnce = */ true>
constexpr auto warp_gemm = []() {
constexpr index_t WarpGemmM = Problem::BlockFmhaShape::Gemm0WarpTile::at(number<0>{});
static_assert(WarpGemmM == 16 || WarpGemmM == 32);
static_assert(WarpGemmM == 4 || WarpGemmM == 16 || WarpGemmM == 32);
if constexpr(std::is_same_v<typename Problem::QDataType, half_t> &&
std::is_same_v<typename Problem::KDataType, half_t> &&
......@@ -113,8 +82,10 @@ struct BlockFmhaPipelineQXCustomPolicy</* QLoadOnce = */ true>
{
if constexpr(WarpGemmM == 32)
return WarpGemmMfmaF16F16F32M32N32K16SwizzleBTransposedCDistribution{};
else // WarpGemmM == 16
else if constexpr(WarpGemmM == 16)
return WarpGemmMfmaF16F16F32M16N16K16TransposedCDistribution{};
else // WarpGemmM == 4
return WarpGemmMfmaF16F16F32M4N64K16{};
}
else if constexpr(std::is_same_v<typename Problem::QDataType, bf16_t> &&
std::is_same_v<typename Problem::KDataType, bf16_t> &&
......@@ -122,8 +93,10 @@ struct BlockFmhaPipelineQXCustomPolicy</* QLoadOnce = */ true>
{
if constexpr(WarpGemmM == 32)
return WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleBTransposedCDistribution{};
else // WarpGemmM == 16
else if constexpr(WarpGemmM == 16)
return WarpGemmMfmaBf16Bf16F32M16N16K16TransposedCDistribution{};
else // WarpGemmM == 4
return WarpGemmMfmaBf16Bf16F32M4N64K16{};
}
else if constexpr(std::is_same_v<typename Problem::QDataType, fp8_t> &&
std::is_same_v<typename Problem::KDataType, fp8_t> &&
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment