Skip to content

GitLab

Commit e43df26a authored by aska-0096's avatar aska-0096
Browse files

temp save, reproduce the v_bfi_b32 issue

parent 9739ede0
Show whitespace changes
Inline Side-by-side
Showing with 146 additions and 92 deletions
include/ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp
View file @ e43df26a
......@@ -280,24 +280,24 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle
const BBlockBuffer& b_block_buf,
CThreadBuffer& c_thread_buf) const
{
// auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
// a_thread_desc_.GetElementSpaceSize());
// auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
// b_thread_desc_.GetElementSpaceSize());
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
b_thread_desc_.GetElementSpaceSize());
StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
FloatAB,
MRepeat,
WmmaK,
true>
a_thread_buf;
// StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
// FloatAB,
// MRepeat,
// WmmaK,
// true>
// a_thread_buf;
StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
FloatAB,
NRepeat,
WmmaK,
true>
b_thread_buf;
// StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
// FloatAB,
// NRepeat,
// WmmaK,
// true>
// b_thread_buf;
static_for<0, KPerBlock / WmmaK, 1>{}([&](auto k) { // k=0,1,2 instead of k=0,kpack*1, ...
static_for<0, MRepeat, 1>{}([&](auto m0) {
......@@ -306,8 +306,8 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle
make_tuple(Number<k * WmmaK / A_K1>{}, m0, I0, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(I0, I0, I0, I0, I0),
a_thread_buf.GetVectorTypeReference(Number<m0*WmmaK>{}).template AsType<FloatAB>());
make_tuple(I0, m0, I0, I0, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
// read B
......@@ -315,28 +315,28 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle
make_tuple(Number<k * WmmaK / B_K1>{}, n0, I0, I0, I0),
b_block_buf,
b_thread_desc_,
make_tuple(I0, I0, I0, I0, I0),
b_thread_buf.GetVectorTypeReference(Number<n0*WmmaK>{}).template AsType<FloatAB>());
// vector_type<FloatAB, WmmaK> a_thread_vec;
// vector_type<FloatAB, WmmaK> b_thread_vec;
// static_for<0, WmmaK, 1>{}([&](auto i) {
// a_thread_vec.template AsType<FloatAB>()(i) =
// a_thread_buf[Number<a_thread_desc_.CalculateOffset(
// make_tuple(i / A_K1, m0, 0, 0, i % A_K1))>{}];
// b_thread_vec.template AsType<FloatAB>()(i) =
// b_thread_buf[Number<b_thread_desc_.CalculateOffset(
// make_tuple(i / B_K1, n0, 0, 0, i % B_K1))>{}];
// });
// using wmma_input_type = typename vector_type<FloatAB, WmmaK>::type;
make_tuple(I0, n0, I0, I0, I0),
b_thread_buf);
vector_type<FloatAB, WmmaK> a_thread_vec;
vector_type<FloatAB, WmmaK> b_thread_vec;
static_for<0, WmmaK, 1>{}([&](auto i) {
a_thread_vec.template AsType<FloatAB>()(i) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(i / A_K1, m0, 0, 0, i % A_K1))>{}];
b_thread_vec.template AsType<FloatAB>()(i) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(i / B_K1, n0, 0, 0, i % B_K1))>{}];
});
using wmma_input_type = typename vector_type<FloatAB, WmmaK>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
wmma_gemm.template Run(
a_thread_buf.GetVectorTypeReference(Number<m0*WmmaK>{}),
b_thread_buf.GetVectorTypeReference(Number<n0*WmmaK>{}),
a_thread_vec.template AsType<wmma_input_type>()(Number<0>{}),
b_thread_vec.template AsType<wmma_input_type>()(Number<0>{}),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
......@@ -346,11 +346,11 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle
protected:
// A[M0, M1, M2, K0 = WmmaK]
static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<WmmaK / A_K1>{}, I1, I1, I1, Number<A_K1>{}));
make_tuple(Number<WmmaK / A_K1>{}, Number<MRepeat>{}, I1, I1, Number<A_K1>{}));
// B[N0, N1, N2, K0 = WmmaK]
static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<WmmaK / B_K1>{}, I1, I1, I1, Number<B_K1>{}));
make_tuple(Number<WmmaK / B_K1>{}, Number<NRepeat>{}, I1, I1, Number<B_K1>{}));
// C[M, N, NumRegWMMA]
static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
......@@ -659,7 +659,7 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_MNKloop
make_tuple(I0, m0, I0, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(I0, I0, I0, I0, I0),
make_tuple(I0, Number<m0>{}, I0, I0, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
......@@ -668,7 +668,7 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_MNKloop
make_tuple(I0, n0, I0, I0, I0),
b_block_buf,
b_thread_desc_,
make_tuple(I0, I0, I0, I0, I0),
make_tuple(I0, Number<n0>{}, I0, I0, I0),
b_thread_buf);
static_for<0, KPerBlock / WmmaK, 1>{}([&](auto k) { // k=0,1,2 instead of k=0,kpack*1, ...
......@@ -678,10 +678,10 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_MNKloop
static_for<0, WmmaK, 1>{}([&](auto i) {
a_thread_vec.template AsType<FloatAB>()(i) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple((k*WmmaK + i) / A_K1, 0, 0, 0, (k*WmmaK + i) % A_K1))>{}];
make_tuple((k*WmmaK + i) / A_K1, m0, 0, 0, (k*WmmaK + i) % A_K1))>{}];
b_thread_vec.template AsType<FloatAB>()(i) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple((k*WmmaK + i) / B_K1, 0, 0, 0, (k*WmmaK + i) % B_K1))>{}];
make_tuple((k*WmmaK + i) / B_K1, n0, 0, 0, (k*WmmaK + i) % B_K1))>{}];
});
using wmma_input_type = typename vector_type<FloatAB, WmmaK>::type;
......@@ -701,11 +701,11 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_MNKloop
protected:
// A[M0, M1, M2, K0 = WmmaK]
static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<KPerBlock / A_K1>{}, I1, I1, I1, Number<A_K1>{}));
make_tuple(Number<KPerBlock / A_K1>{}, Number<MRepeat>{}, I1, I1, Number<A_K1>{}));
// B[N0, N1, N2, K0 = WmmaK]
static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<KPerBlock / B_K1>{}, I1, I1, I1, Number<B_K1>{}));
make_tuple(Number<KPerBlock / B_K1>{}, Number<NRepeat>{}, I1, I1, Number<B_K1>{}));
// C[M, N, NumRegWMMA]
static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
......@@ -716,7 +716,7 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_MNKloop
decltype(a_block_desc_k0_m0_m1_m2_k1),
decltype(a_thread_desc_),
Sequence<KPerBlock / A_K1, 1, 1, 1, A_K1>,
Sequence<3, 0, 1, 2, 4>,
Sequence<0, 1, 2, 3, 4>,
4,
A_K1,
A_K1>;
......@@ -726,7 +726,7 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_MNKloop
decltype(b_block_desc_k0_n0_n1_n2_k1),
decltype(b_thread_desc_),
Sequence<KPerBlock / B_K1, 1, 1, 1, B_K1>,
Sequence<3, 0, 1, 2, 4>,
Sequence<0, 1, 2, 3, 4>,
4,
B_K1,
B_K1>;
......@@ -1012,6 +1012,14 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_FIFO
static_for<0, KPerBlock, WmmaK>{}([&](auto iWmmaK){
static_for<0, NRepeat, 1>{}([&](auto iN){
b_thread_copy_.Run(b_block_desc_k0_n0_n1_n2_k1,
make_tuple(Number<iWmmaK/B_K1>{}, Number<iN>{}, I0, I0, I0),
b_block_buf,
b_thread_desc_,
make_tuple(I0, Number<iN>{}, I0, I0, I0),
b_thread_buf);
});
// Stage 1: Cut to Repeat Retangle to Square, assume MRepeat > NRepeat
static_for<0, RepeatDiff, 1>{}([&](auto iCut){
a_thread_copy_.Run(a_block_desc_k0_m0_m1_m2_k1,
......@@ -1021,12 +1029,12 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_FIFO
make_tuple(I0, Number<iCut>{}, I0, I0, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto iN){
b_thread_copy_.Run(b_block_desc_k0_n0_n1_n2_k1,
make_tuple(Number<iWmmaK/B_K1>{}, Number<iN>{}, I0, I0, I0),
b_block_buf,
b_thread_desc_,
make_tuple(I0, Number<iN>{}, I0, I0, I0),
b_thread_buf);
// b_thread_copy_.Run(b_block_desc_k0_n0_n1_n2_k1,
// make_tuple(Number<iWmmaK/B_K1>{}, Number<iN>{}, I0, I0, I0),
// b_block_buf,
// b_thread_desc_,
// make_tuple(I0, Number<iN>{}, I0, I0, I0),
// b_thread_buf);
vector_type<FloatAB, WmmaK> a_thread_vec;
vector_type<FloatAB, WmmaK> b_thread_vec;
......@@ -1042,30 +1050,34 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_FIFO
using wmma_input_type = typename vector_type<FloatAB, WmmaK>::type;
constexpr index_t c_offset = c_thread_desc_.CalculateOffset(make_tuple(iCut, iN, 0));
s_nop();
wmma_gemm.template Run(
a_thread_vec.template AsType<wmma_input_type>()(Number<0>{}),
b_thread_vec.template AsType<wmma_input_type>()(Number<0>{}),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
s_nop();
});
});
// Stage 2: Run FIFO fashion loopover in Square
static_for<0, NRepeat, 1>{}([&](auto WmmaInnerloop){
// Row Repeatation
static_for<WmmaInnerloop, NRepeat, 1>{}([&](auto iN){
a_thread_copy_.Run(a_block_desc_k0_m0_m1_m2_k1,
make_tuple(Number<iWmmaK/A_K1>{}, Number<WmmaInnerloop+RepeatDiff>{}, I0, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(I0, Number<WmmaInnerloop+RepeatDiff>{}, I0, I0, I0),
a_thread_buf);
b_thread_copy_.Run(b_block_desc_k0_n0_n1_n2_k1,
make_tuple(Number<iWmmaK/B_K1>{}, Number<iN>{}, I0, I0, I0),
b_block_buf,
b_thread_desc_,
make_tuple(I0, Number<iN>{}, I0, I0, I0),
b_thread_buf);
});
// Stage 2: Run FIFO fashion loopover in Square
static_for<0, NRepeat, 1>{}([&](auto WmmaInnerloop){
// Row Repeatation
static_for<WmmaInnerloop, NRepeat, 1>{}([&](auto iN){
// b_thread_copy_.Run(b_block_desc_k0_n0_n1_n2_k1,
// make_tuple(Number<iWmmaK/B_K1>{}, Number<iN>{}, I0, I0, I0),
// b_block_buf,
// b_thread_desc_,
// make_tuple(I0, Number<iN>{}, I0, I0, I0),
// b_thread_buf);
vector_type<FloatAB, WmmaK> a_thread_vec;
vector_type<FloatAB, WmmaK> b_thread_vec;
......@@ -1081,27 +1093,29 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_FIFO
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(WmmaInnerloop+RepeatDiff, iN, 0));
s_nop();
wmma_gemm.template Run(
a_thread_vec.template AsType<wmma_input_type>()(Number<0>{}),
b_thread_vec.template AsType<wmma_input_type>()(Number<0>{}),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
s_nop();
});
// WmmaInnerloop++
// Col Repeatation
static_for<WmmaInnerloop+1+RepeatDiff, MRepeat, 1>{}([&](auto iM){
a_thread_copy_.Run(a_block_desc_k0_m0_m1_m2_k1,
make_tuple(Number<iWmmaK/A_K1>{}, Number<iM>{}, I0, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(I0, Number<iM>{}, I0, I0, I0),
a_thread_buf);
b_thread_copy_.Run(b_block_desc_k0_n0_n1_n2_k1,
make_tuple(Number<iWmmaK/B_K1>{}, Number<WmmaInnerloop>{}, I0, I0, I0),
b_block_buf,
b_thread_desc_,
make_tuple(I0, Number<WmmaInnerloop>{}, I0, I0, I0),
b_thread_buf);
// a_thread_copy_.Run(a_block_desc_k0_m0_m1_m2_k1,
// make_tuple(Number<iWmmaK/A_K1>{}, Number<iM>{}, I0, I0, I0),
// a_block_buf,
// a_thread_desc_,
// make_tuple(I0, Number<iM>{}, I0, I0, I0),
// a_thread_buf);
// b_thread_copy_.Run(b_block_desc_k0_n0_n1_n2_k1,
// make_tuple(Number<iWmmaK/B_K1>{}, Number<WmmaInnerloop>{}, I0, I0, I0),
// b_block_buf,
// b_thread_desc_,
// make_tuple(I0, Number<WmmaInnerloop>{}, I0, I0, I0),
// b_thread_buf);
vector_type<FloatAB, WmmaK> a_thread_vec;
vector_type<FloatAB, WmmaK> b_thread_vec;
......@@ -1117,10 +1131,12 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_FIFO
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(iM, WmmaInnerloop, 0));
s_nop();
wmma_gemm.template Run(
a_thread_vec.template AsType<wmma_input_type>()(Number<0>{}),
b_thread_vec.template AsType<wmma_input_type>()(Number<0>{}),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
s_nop();
});
});
});
......@@ -1144,7 +1160,7 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_FIFO
decltype(a_block_desc_k0_m0_m1_m2_k1),
decltype(a_thread_desc_),
Sequence<WmmaK / A_K1, 1, 1, 1, A_K1>,
Sequence<3, 0, 1, 2, 4>,
Sequence<0, 1, 2, 3, 4>,
4,
A_K1,
A_K1>;
......@@ -1154,7 +1170,7 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_FIFO
decltype(b_block_desc_k0_n0_n1_n2_k1),
decltype(b_thread_desc_),
Sequence<WmmaK / B_K1, 1, 1, 1, B_K1>,
Sequence<3, 0, 1, 2, 4>,
Sequence<0, 1, 2, 3, 4>,
4,
B_K1,
B_K1>;
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp
View file @ e43df26a
......@@ -310,7 +310,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_wmma
constexpr auto WmmaK = 16;
constexpr auto KPack = math::integer_least_multiple(K1, WmmaK);
using BlockwiseGemm = BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle<
using BlockwiseGemm = BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_FIFO<
BlockSize,
FloatAB,
FloatAcc,
......@@ -367,7 +367,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_wmma
constexpr auto WmmaK = 16;
constexpr auto KPack = math::integer_least_multiple(K1, WmmaK);
using BlockwiseGemm = BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle<
using BlockwiseGemm = BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_FIFO<
BlockSize,
FloatAB,
FloatAcc,
......@@ -540,7 +540,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_wmma
constexpr auto KPack = math::integer_least_multiple(K1, WmmaK);
auto blockwise_gemm =
BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle<BlockSize,
BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_FIFO<BlockSize,
FloatAB,
FloatAcc,
decltype(a_block_desc_k0perblock_mperblock_k1),
......
include/ck/utility/amd_inline_asm.hpp
View file @ e43df26a
......@@ -360,9 +360,7 @@ __device__ void amd_assembly_wmma_f32_16x16x16_f16_w32(half16_t a,
half16_t b,
float8_t& c)
{
asm volatile("\n \
v_wmma_f32_16x16x16_f16_w32 %0, %1, %2, %0\n \
"
asm volatile("v_wmma_f32_16x16x16_f16 %0, %1, %2, %0"
: "=v"(c)
: "v"(a), "v"(b), "0"(c));
}
......
include/ck/utility/amd_wmma.hpp
View file @ e43df26a
......@@ -4,6 +4,7 @@
#ifndef CK_AMD_WMMA_HPP
#define CK_AMD_WMMA_HPP
#include "ck/utility/amd_inline_asm.hpp"
#include "data_type.hpp"
// TODO: Add arch limitation
namespace ck {
......@@ -20,8 +21,10 @@ struct intrin_wmma_f32_16x16x16_f16_w32<16, 16>
template <class FloatC>
__device__ static void Run(const half16_t& reg_a, const half16_t& reg_b, FloatC& reg_c)
{
reg_c.template AsType<float8_t>()(Number<0>{}) = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32(
reg_a, reg_b, reg_c.template AsType<float8_t>()[Number<0>{}]);
// * Inline assembly need to elimate the duplicated data load, compiler won't help you delete them.
amd_assembly_wmma_f32_16x16x16_f16_w32(reg_a, reg_b, reg_c.template AsType<float8_t>()(Number<0>{}));
// reg_c.template AsType<float8_t>()(Number<0>{}) = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32(
// reg_a, reg_b, reg_c.template AsType<float8_t>()[Number<0>{}]);
}
};
......
test/wmma_op/wmma_op_util.hpp
View file @ e43df26a
......@@ -97,6 +97,7 @@ builtin_wmma_naive_selector<int4x16_t,
template <typename src_t, typename dst_t, typename acc_t, index_t acc_num>
__global__ void matmul(const src_t* a, const src_t* b, dst_t* c)
{
__shared__ src_t p_shared[16*16*2];
const int lIdx = threadIdx.x;
// a and b fragments are stored in 8 VGPRs each, in packed format, so 16 elements each for a and
// b a_frag will store one column of the 16x16 matrix tile b_frag will store one row of the
......@@ -104,6 +105,9 @@ __global__ void matmul(const src_t* a, const src_t* b, dst_t* c)
using src_vec = typename vector_type<src_t, 16>::type;
src_vec a_frag = {};
src_vec b_frag = {};
src_vec a_temp = {};
src_vec b_temp = {};
// initialize c fragment to 0
using acc_vec = StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, acc_t, 1, acc_num, true>;
acc_vec c_thread_buf_;
......@@ -112,19 +116,52 @@ __global__ void matmul(const src_t* a, const src_t* b, dst_t* c)
// see https://atlvsp3.amd.com/sp3_gfx11_5_instructions.pdf page 482
// TODO: remove this dependency in gfx12 https://ontrack-internal.amd.com/browse/DEGFXSP3-101
const int lane = lIdx % 16;
const int lane_lo = lIdx / 2;
const int lane_hi = lIdx % 2;
for(int ele = 0; ele < 8; ++ele)
{
a_temp[ele] = a[8 * lane_hi + 16 * lane_lo + ele];
}
for(int ele = 0; ele < 8; ++ele)
{
b_temp[ele] = b[8 * lane_hi + 16 * lane_lo + ele];
}
__syncthreads();
for(int ele = 0; ele < 8; ++ele)
{
p_shared[8*16*lane_hi + 8 * lane_lo + ele] = a_temp[ele];
}
for(int ele = 0; ele < 8; ++ele)
{
p_shared[8*16*lane_hi + 8 * lane_lo + ele + 16*16] = b_temp[ele];
}
asm volatile("\
s_waitcnt lgkmcnt(0) \n \
s_barrier \
" ::);
for(int ele = 0; ele < 16; ++ele)
{
b_frag[ele] = b[16 * lane + ele];
b_frag[ele] = p_shared[(ele/8) * 16*8 + 8 * lane + ele%8 + 16*16];
}
// follow origin design
for(int ele = 0; ele < 16; ++ele)
{
a_frag[ele] = a[16 * lane + ele];
a_frag[ele] = p_shared[(ele/8) * 16*8 + 8 * lane + ele%8];
}
asm volatile("\
s_waitcnt lgkmcnt(0) \n \
s_barrier \
" ::);
// sync threads, similar to mma_sync
__syncthreads();
// __syncthreads();
builtin_wmma_naive_selector<src_vec, acc_vec>(a_frag, b_frag, c_thread_buf_);
__syncthreads();
// wait for results, similar to mma_sync
......
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