Skip to content

GitLab

"tests/layers/__init__.py" did not exist on "fd158e88e82c3fa848017c62a7eccb49a5c64f78"
Unverified Commit d46196f2 authored by Adam Osewski's avatar Adam Osewski Committed by GitHub
Browse files

Enhance printing functionality (#1751) 



* Added object print with all template parameters

* fix clang format

---------
Co-authored-by: default avatarravil-mobile <ravil.aviva.com@gmail.com>
Co-authored-by: default avatarillsilin <Illia.Silin@amd.com>
parent 0fd6978d
Show whitespace changes
Inline Side-by-side
Showing with 2095 additions and 1787 deletions
include/ck/tensor_operation/gpu/device/device_base.hpp
View file @ d46196f2
......@@ -5,6 +5,8 @@
#include <string>
#include <sstream>
#include <regex>
#include <optional>
#include "ck/stream_config.hpp"
......@@ -12,6 +14,34 @@ namespace ck {
namespace tensor_operation {
namespace device {
#define GET_OBJECT_NAME_IMLP \
std::optional<std::string> GetObjectName() const override \
{ \
std::string str = __PRETTY_FUNCTION__; \
static std::regex obj_name_expr{"<std::string> (.*)::GetObjectName"}; \
std::smatch match; \
if(!std::regex_search(str, match, obj_name_expr)) \
{ \
return str; \
} \
return std::string(match[1]) + ';'; \
}
#define GET_TEMPLATE_INFO_IMPL \
std::optional<std::string> GetTemplateInfo() const override \
{ \
std::string str = __PRETTY_FUNCTION__; \
static std::regex template_expr{"\\[(.*)\\]"}; \
std::smatch match; \
if(!std::regex_search(str, match, template_expr)) \
{ \
return std::nullopt; \
} \
return std::string(match[1]); \
}
#define REGISTER_EXTRA_PRINTING_METHODS GET_OBJECT_NAME_IMLP GET_TEMPLATE_INFO_IMPL
struct BaseArgument
{
BaseArgument() = default;
......@@ -48,6 +78,10 @@ struct BaseOperator
virtual std::string GetTypeIdName() const { return typeid(*this).name(); }
virtual std::optional<std::string> GetObjectName() const { return std::nullopt; }
virtual std::optional<std::string> GetTemplateInfo() const { return std::nullopt; }
virtual std::string GetTypeIdHashCode() const
{
std::ostringstream oss;
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp
View file @ d46196f2
......@@ -729,6 +729,7 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
return str.str();
}
REGISTER_EXTRA_PRINTING_METHODS
};
} // namespace device
......
include/ck_tile/ops/flatmm/block/uk/flatmm_sn_uk_gfx9_32x128x512_1x4x1_16x16x16.inc
View file @ d46196f2
......@@ -3,9 +3,9 @@
#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_) \
#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" \
......@@ -14,599 +14,804 @@
" 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_) \
#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 @ d46196f2
......@@ -3,9 +3,9 @@
#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_) \
#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" \
......@@ -14,695 +14,756 @@
" 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_) \
#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"
" 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"
" 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_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_PK_CVT_
#undef _UK_ATOMIC_ADD_
include/ck_tile/ops/flatmm/block/uk/flatmm_uk_gfx9_32x512x128_1x1x1_16x16x16.inc
View file @ d46196f2
......@@ -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_
profiler/include/profiler/profile_gemm_universal_impl.hpp
View file @ d46196f2
......@@ -144,6 +144,7 @@ bool profile_gemm_universal_impl(int do_verification,
}
std::string best_op_name;
std::optional<std::string> best_op_object_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
......@@ -226,6 +227,7 @@ bool profile_gemm_universal_impl(int do_verification,
}
std::string op_name = op_ptr->GetTypeString();
std::optional<std::string> op_obj_name = op_ptr->GetObjectName();
float ave_time = invoker_ptr->Run(argument_ptr.get(),
StreamConfig{nullptr,
......@@ -252,6 +254,7 @@ bool profile_gemm_universal_impl(int do_verification,
if(tflops > best_tflops && ave_time > 1e-10)
{
best_op_name = op_name;
best_op_object_name = op_obj_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
......@@ -306,6 +309,9 @@ bool profile_gemm_universal_impl(int do_verification,
<< " : " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
<< " GB/s, " << best_op_name << std::endl;
if(best_op_object_name)
std::cout << best_op_object_name.value() << std::endl;
return pass;
}
......
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