summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rwxr-xr-xcuda-kernels/Makefile4
-rw-r--r--cuda-kernels/tensor_core.cu57
-rw-r--r--cuda-kernels/tensorcore_type32_32.cu202
-rw-r--r--src/cuda-sim/instructions.cc110
4 files changed, 309 insertions, 64 deletions
diff --git a/cuda-kernels/Makefile b/cuda-kernels/Makefile
index 673460f..73a4f0c 100755
--- a/cuda-kernels/Makefile
+++ b/cuda-kernels/Makefile
@@ -1,5 +1,5 @@
-all: tensor_core.cu
- nvcc --gpu-architecture=compute_70 --gpu-code=compute_70 -lcudart -g -o tensor_core tensor_core.cu
+all: tensorcore_type32_32.cu
+ nvcc --gpu-architecture=compute_70 --gpu-code=compute_70 -lcudart -g -o tensor_core tensorcore_type32_32.cu
# nvcc -arch=sm_70 -lcudart -g -o tensor_core tensor_core.cu
.PHONY:
diff --git a/cuda-kernels/tensor_core.cu b/cuda-kernels/tensor_core.cu
index d5c1b40..b7090c4 100644
--- a/cuda-kernels/tensor_core.cu
+++ b/cuda-kernels/tensor_core.cu
@@ -132,11 +132,7 @@ int main(int argc, char* argv[]) {
cudaErrCheck(cudaEventCreate(&startWMMA));
cudaErrCheck(cudaEventCreate(&stopWMMA));
-
-
-
// Use tensor cores
-
cudaErrCheck(cudaMalloc((void**)&a_fp32, MATRIX_M * MATRIX_K * sizeof(float)));
cudaErrCheck(cudaMalloc((void**)&b_fp32, MATRIX_K * MATRIX_N * sizeof(float)));
@@ -151,8 +147,6 @@ int main(int argc, char* argv[]) {
a_host_wmma = (float*)malloc(MATRIX_M * MATRIX_K * sizeof(float));
b_host_wmma = (float*)malloc(MATRIX_K * MATRIX_N * sizeof(float));
-
-
// printf("a_fp32\n");
for(int m=0;m<MATRIX_M;m++){
for(int n=0;n<MATRIX_K;n++){
@@ -160,13 +154,14 @@ int main(int argc, char* argv[]) {
}
//printf(";\n");
}
+
// printf("b_fp32\n");
for(int m=0;m<MATRIX_K;m++){
for(int n=0;n<MATRIX_N;n++){
b_host_wmma[m*MATRIX_N+n]=(m*MATRIX_N+n)%10;
-// printf("%f ",b_host_wmma[m*MATRIX_N+n]);
+ // printf("%f ",b_host_wmma[m*MATRIX_N+n]);
}
-// printf(";\n");
+ // printf(";\n");
}
cudaErrCheck(cudaMemcpy(a_fp32,a_host_wmma, MATRIX_M * MATRIX_K * sizeof(float), cudaMemcpyHostToDevice));
cudaErrCheck(cudaMemcpy(b_fp32,b_host_wmma, MATRIX_K * MATRIX_N * sizeof(float), cudaMemcpyHostToDevice));
@@ -215,36 +210,36 @@ int main(int argc, char* argv[]) {
cudaErrCheck(cudaMemcpy(c_host_wmma, c_wmma, MATRIX_M * MATRIX_N * sizeof(float), cudaMemcpyDeviceToHost));
// printf("c_host\n");
// for(int m=0;m<MATRIX_M;m++){
-// for(int n=0;n<MATRIX_N;n++){
-// printf("%f ",c_host_wmma[m*MATRIX_N+n]);
-// }
-// printf(";\n");
+ //for(int n=0;n<MATRIX_N;n++){
+ //printf("%f ",c_host_wmma[m*MATRIX_N+n]);
+ //}
+ //printf(";\n");
// }
- float wmmaTime;
- cudaErrCheck(cudaEventSynchronize(stopWMMA));
- cudaErrCheck(cudaEventElapsedTime(&wmmaTime, startWMMA, stopWMMA));
- printf("wmma took %fms\n", wmmaTime);
- //printf("Clock=%d",stopWMMA-startWMMA);
- printf("\nFor a faster code using wmma you should check out the cudaTensorCoreGemm sample in the CUDA Toolkit.\nThis code was written as a demo only!\n\n");
+ float wmmaTime;
+ cudaErrCheck(cudaEventSynchronize(stopWMMA));
+ cudaErrCheck(cudaEventElapsedTime(&wmmaTime, startWMMA, stopWMMA));
+ printf("wmma took %fms\n", wmmaTime);
+ //printf("Clock=%d",stopWMMA-startWMMA);
+ printf("\nFor a faster code using wmma you should check out the cudaTensorCoreGemm sample in the CUDA Toolkit.\nThis code was written as a demo only!\n\n");
- cudaErrCheck(cudaEventDestroy(startWMMA));
- cudaErrCheck(cudaEventDestroy(stopWMMA));
+ cudaErrCheck(cudaEventDestroy(startWMMA));
+ cudaErrCheck(cudaEventDestroy(stopWMMA));
-
- cudaErrCheck(cudaFree(a_fp32));
- cudaErrCheck(cudaFree(b_fp32));
- cudaErrCheck(cudaFree(a_fp16));
- cudaErrCheck(cudaFree(b_fp16));
+
+ cudaErrCheck(cudaFree(a_fp32));
+ cudaErrCheck(cudaFree(b_fp32));
+ cudaErrCheck(cudaFree(a_fp16));
+ cudaErrCheck(cudaFree(b_fp16));
- cudaErrCheck(cudaFree(c));
- cudaErrCheck(cudaFree(c_wmma));
-
- free(c_host_wmma);
+ cudaErrCheck(cudaFree(c));
+ cudaErrCheck(cudaFree(c_wmma));
+
+ free(c_host_wmma);
- cudaErrCheck(cudaDeviceReset());
- return 0;
+ cudaErrCheck(cudaDeviceReset());
+ return 0;
}
diff --git a/cuda-kernels/tensorcore_type32_32.cu b/cuda-kernels/tensorcore_type32_32.cu
new file mode 100644
index 0000000..0d26163
--- /dev/null
+++ b/cuda-kernels/tensorcore_type32_32.cu
@@ -0,0 +1,202 @@
+#include <stdio.h>
+#include <curand.h>
+
+// Define some error checking macros.
+#define cudaErrCheck(stat) { cudaErrCheck_((stat), __FILE__, __LINE__); }
+void cudaErrCheck_(cudaError_t stat, const char *file, int line) {
+ if (stat != cudaSuccess) {
+ fprintf(stderr, "CUDA Error: %s %s %d\n", cudaGetErrorString(stat), file, line);
+ }
+}
+
+#define curandErrCheck(stat) { curandErrCheck_((stat), __FILE__, __LINE__); }
+void curandErrCheck_(curandStatus_t stat, const char *file, int line) {
+ if (stat != CURAND_STATUS_SUCCESS) {
+ fprintf(stderr, "cuRand Error: %d %s %d\n", stat, file, line);
+ }
+}
+
+#include <mma.h>
+using namespace nvcuda;
+
+// Must be multiples of 16 for wmma code to work
+#define MATRIX_M (16)
+#define MATRIX_N (16)
+#define MATRIX_K (16)
+
+
+// The only dimensions currently supported by WMMA
+const int WMMA_M = 16;
+const int WMMA_N = 16;
+const int WMMA_K = 16;
+
+__global__ void wmma_example(half *a, half *b, float *c,float *d_fp16, int M, int N, int K) {
+ //unsigned int start_time=0,end_time=0;
+ //start_time=clock();
+
+ // Declare the fragments
+ wmma::fragment<wmma::matrix_a, WMMA_M, WMMA_N, WMMA_K, half, wmma::col_major> a_frag;
+ wmma::fragment<wmma::matrix_b, WMMA_M, WMMA_N, WMMA_K, half, wmma::col_major> b_frag;
+ wmma::fragment<wmma::accumulator, WMMA_M, WMMA_N, WMMA_K, float> c_frag;
+
+ // Bounds checking
+ wmma::load_matrix_sync(a_frag, a, K);
+ wmma::load_matrix_sync(b_frag, b, K);
+ wmma::load_matrix_sync(c_frag, c, N,wmma::mem_col_major);
+ wmma::mma_sync(c_frag, a_frag, b_frag, c_frag);
+
+ wmma::store_matrix_sync(d_fp16, c_frag, N, wmma::mem_col_major);
+ //printf("clock=%d",end_time-start_time);
+}
+
+__global__ void convertFp32ToFp16 (half *out, float *in, int n) {
+ int idx = blockDim.x * blockIdx.x + threadIdx.x;
+ if (idx < n) {
+ out[idx] = in[idx];
+ }
+}
+__global__ void convertFp16ToFp32 (float *out, half *in, int n) {
+ int idx = blockDim.x * blockIdx.x + threadIdx.x;
+ if (idx < n) {
+ out[idx] = in[idx];
+ }
+}
+
+int main(int argc, char* argv[]) {
+ float *a_fp32;
+ float *b_fp32;
+ float *c_fp32;
+ float *d_fp32;
+
+ half *a_fp16;
+ half *b_fp16;
+ // half *c_fp16;
+ // half *d_fp16;
+
+ float *a_host_wmma;
+ float *b_host_wmma;
+ float *c_host_wmma;
+ float *d_host_wmma;
+ float *d_cal_host_wmma;
+
+ cudaEvent_t startWMMA;
+ cudaEvent_t stopWMMA;
+
+
+ cudaErrCheck(cudaEventCreate(&startWMMA));
+ cudaErrCheck(cudaEventCreate(&stopWMMA));
+
+ // Use tensor cores
+ cudaErrCheck(cudaMalloc((void**)&a_fp32, MATRIX_M * MATRIX_K * sizeof(float)));
+ cudaErrCheck(cudaMalloc((void**)&b_fp32, MATRIX_K * MATRIX_N * sizeof(float)));
+ cudaErrCheck(cudaMalloc((void**)&c_fp32, MATRIX_K * MATRIX_N * sizeof(float)));
+ cudaErrCheck(cudaMalloc((void**)&d_fp32, MATRIX_K * MATRIX_N * sizeof(float)));
+ cudaErrCheck(cudaMalloc((void**)&a_fp16, MATRIX_M * MATRIX_K * sizeof(half)));
+ cudaErrCheck(cudaMalloc((void**)&b_fp16, MATRIX_K * MATRIX_N * sizeof(half)));
+ //cudaErrCheck(cudaMalloc((void**)&c_fp16, MATRIX_K * MATRIX_N * sizeof(half)));
+ //cudaErrCheck(cudaMalloc((void**)&d_fp16, MATRIX_K * MATRIX_N * sizeof(half)));
+
+
+ a_host_wmma = (float*)malloc(MATRIX_M * MATRIX_K * sizeof(float));
+ b_host_wmma = (float*)malloc(MATRIX_K * MATRIX_N * sizeof(float));
+ c_host_wmma = (float*)malloc(MATRIX_M * MATRIX_N * sizeof(float));
+ d_host_wmma = (float*)malloc(MATRIX_M * MATRIX_N * sizeof(float));
+ d_cal_host_wmma = (float*)malloc(MATRIX_M * MATRIX_N * sizeof(float));
+
+ //printf("a_fp32\n");
+ for(int m=0;m<MATRIX_M;m++){
+ for(int n=0;n<MATRIX_K;n++){
+ a_host_wmma[m*MATRIX_K+n]=(m*MATRIX_K+n)%10;
+ // printf("%f ",a_host_wmma[m*MATRIX_K+n]);
+ }
+ //printf(";\n");
+ }
+
+ //printf("b_fp32\n");
+ for(int m=0;m<MATRIX_K;m++){
+ for(int n=0;n<MATRIX_N;n++){
+ b_host_wmma[m*MATRIX_N+n]=(m*MATRIX_N+n)%10;
+ // printf("%f ",b_host_wmma[m*MATRIX_N+n]);
+ }
+ // printf(";\n");
+ }
+
+ //printf("c_fp32\n");
+ for(int m=0;m<MATRIX_M;m++){
+ for(int n=0;n<MATRIX_N;n++){
+ c_host_wmma[m*MATRIX_N+n]=(m*MATRIX_N+n)%10;
+ d_cal_host_wmma[m*MATRIX_N+n]=0;
+ // printf("%f ",c_host_wmma[m*MATRIX_N+n]);
+ }
+ }
+ for(int m=0;m<MATRIX_M;m++){
+ for(int n=0;n<MATRIX_N;n++){
+ for(int k=0;k<MATRIX_K;k++){
+ d_cal_host_wmma[m*MATRIX_N+n]+= a_host_wmma[m*MATRIX_K+k]*b_host_wmma[k*MATRIX_K+n];
+ }
+ d_cal_host_wmma[m*MATRIX_N+n]+=c_host_wmma[m*MATRIX_N+n];
+ }
+ }
+
+
+ cudaErrCheck(cudaMemcpy(a_fp32,a_host_wmma, MATRIX_M * MATRIX_K * sizeof(float), cudaMemcpyHostToDevice));
+ cudaErrCheck(cudaMemcpy(b_fp32,b_host_wmma, MATRIX_K * MATRIX_N * sizeof(float), cudaMemcpyHostToDevice));
+ cudaErrCheck(cudaMemcpy(c_fp32,c_host_wmma, MATRIX_M * MATRIX_N * sizeof(float), cudaMemcpyHostToDevice));
+
+ convertFp32ToFp16 <<< (MATRIX_M * MATRIX_K + 255) / 256, 256 >>> (a_fp16, a_fp32, MATRIX_M * MATRIX_K);
+ convertFp32ToFp16 <<< (MATRIX_K * MATRIX_N + 255) / 256, 256 >>> (b_fp16, b_fp32, MATRIX_K * MATRIX_N);
+ //convertFp32ToFp16 <<< (MATRIX_M * MATRIX_N + 255) / 256, 256 >>> (c_fp16, c_fp32, MATRIX_K * MATRIX_N);
+
+ printf("\nM = %d, N = %d, K = %d. \n", MATRIX_M, MATRIX_N, MATRIX_K);
+
+ printf("Running with wmma...\n");
+ cudaErrCheck(cudaEventRecord(startWMMA));
+ wmma_example <<< 1, 32>>> (a_fp16, b_fp16, c_fp32, d_fp32 , MATRIX_M, MATRIX_N, MATRIX_K);
+ cudaErrCheck(cudaEventRecord(stopWMMA));
+ cudaErrCheck(cudaEventSynchronize(stopWMMA));
+
+ // Error checking
+ printf("\nChecking results...\n");
+ cudaErrCheck(cudaMemcpy(d_host_wmma, d_fp32, MATRIX_M * MATRIX_N * sizeof(float), cudaMemcpyDeviceToHost));
+
+ printf("Results verified: cublas and WMMA agree.\n\n");
+ float wmmaTime;
+ cudaErrCheck(cudaEventElapsedTime(&wmmaTime, startWMMA, stopWMMA));
+ printf("wmma took %fms\n", wmmaTime);
+
+ cudaErrCheck(cudaEventDestroy(startWMMA));
+ cudaErrCheck(cudaEventDestroy(stopWMMA));
+
+ printf("D_CALCULATED\n");
+ for(int m=0;m<MATRIX_M;m++){
+ for(int n=0;n<MATRIX_N;n++){
+ printf("%.2f,",d_cal_host_wmma[m*MATRIX_N+n]);
+ }
+ printf("\n");
+ }
+ printf("D_WMMA\n");
+ for(int m=0;m<MATRIX_M;m++){
+ for(int n=0;n<MATRIX_N;n++){
+ printf("%.2f,",d_host_wmma[m*MATRIX_N+n]);
+ }
+ printf("\n");
+ }
+
+ cudaErrCheck(cudaFree(a_fp32));
+ cudaErrCheck(cudaFree(b_fp32));
+ cudaErrCheck(cudaFree(c_fp32));
+ cudaErrCheck(cudaFree(d_fp32));
+ cudaErrCheck(cudaFree(a_fp16));
+ cudaErrCheck(cudaFree(b_fp16));
+ //cudaErrCheck(cudaFree(c_fp16));
+ //cudaErrCheck(cudaFree(d_fp16));
+
+ free(a_host_wmma);
+ free(b_host_wmma);
+ free(c_host_wmma);
+ free(d_host_wmma);
+ cudaErrCheck(cudaDeviceReset());
+ return 0;
+}
+
+
diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc
index 6f5abaa..70aee35 100644
--- a/src/cuda-sim/instructions.cc
+++ b/src/cuda-sim/instructions.cc
@@ -669,7 +669,7 @@ void ptx_thread_info::set_wmma_vector_operand_values( const operand_info &dst,
printf("error:set_wmma_vector_operands");
}
- m_last_set_operand_value = data1;
+ m_last_set_operand_value = data8;
}
#define my_abs(a) (((a)<0)?(-a):(a))
@@ -1535,6 +1535,7 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
ptx_thread_info *thread;
unsigned type = pI->get_type();
+ unsigned type2 = pI->get_type2();
int tid = inst.warp_id_func() * core->get_warp_size();
const operand_info &dst = pI->operand_lookup(0);
@@ -1553,6 +1554,7 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
unsigned nelem = src_a.get_vect_nelem();
ptx_reg_t v[8];
thread->get_vector_operand_values( src_a, v, nelem );
+ if(i!=3||((i==3)&&(type==F16_TYPE))){
printf("%x ",v[0].f16);
printf("%x ",v[1].f16);
printf("%x ",v[2].f16);
@@ -1561,6 +1563,18 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
printf("%x ",v[5].f16);
printf("%x ",v[6].f16);
printf("%x ",v[7].f16);
+ }
+ else{
+ printf("%f ",v[0].f32);
+ printf("%f ",v[1].f32);
+ printf("%f ",v[2].f32);
+ printf("%f ",v[3].f32);
+ printf("%f ",v[4].f32);
+ printf("%f ",v[5].f32);
+ printf("%f ",v[6].f32);
+ printf("%f ",v[7].f32);
+ }
+
switch(i) {
case 1 ://operand 1
for(k=0;k<8;k++)
@@ -1587,7 +1601,7 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
printf("%x ",matrix_a[i][j].f16);
}
printf("\n");
- }
+ }
printf("MATRIX_B\n");
for (i=0;i<16;i++){
for(j=0;j<16;j++){
@@ -1598,7 +1612,10 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
printf("MATRIX_C\n");
for (i=0;i<16;i++){
for(j=0;j<16;j++){
- printf("%x ",matrix_c[i][j].f16);
+ if(type==F16_TYPE)
+ printf("%x ",matrix_c[i][j].f16);
+ else
+ printf("%f ",matrix_c[i][j].f32);
}
printf("\n");
}
@@ -1606,6 +1623,8 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
for(j=0;j<16;j++){
matrix_d[i][j].f16=0;
}
+ }
+
printf("MATRIX_D\n");
for (i=0;i<16;i++){
for(j=0;j<16;j++){
@@ -1613,20 +1632,28 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
}
printf("\n");
}
- }
-
+ float temp;
for (i=0;i<16;i++){
for(j=0;j<16;j++){
for(k=0;k<16;k++){
- matrix_d[i][j].f16=matrix_d[i][j].f16+matrix_a[i][k].f16*matrix_b[j][k].f16;
+ matrix_d[i][j].f16=matrix_d[i][j].f16+matrix_a[i][k].f16*matrix_b[k][j].f16;
+ }
+ if(type==F16_TYPE)
+ matrix_d[i][j].f16+=matrix_c[i][j].f16;
+ else{
+ temp=matrix_d[i][j].f16;
+ temp+=matrix_c[i][j].f32;
+ matrix_d[i][j].f32=temp;
}
- matrix_d[i][j].f16+=matrix_c[i][j].f16;
}
}
printf("MATRIX_D\n");
for (i=0;i<16;i++){
for(j=0;j<16;j++){
- printf("%x ",matrix_d[i][j].f16);
+ if(type==F16_TYPE)
+ printf("%x ",matrix_d[i][j].f16);
+ else
+ printf("%.2f ",matrix_d[i][j].f32);
}
printf("\n");
}
@@ -2650,16 +2677,16 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
{
size_t size;
int t;
- int thrd,odd,inx;
+ int thrd,odd,inx,k;
ptx_thread_info *thread;
const operand_info &src = pI->operand_lookup(1);
const operand_info &src1 = pI->operand_lookup(0);
const operand_info &src2 = pI->operand_lookup(2);
int tid = inst.warp_id_func()*core->get_warp_size();
+ unsigned type = pI->get_type();
for (thrd=0; thrd < core->get_warp_size(); thrd++) {
- unsigned type = pI->get_type();
thread = core->get_thread_info()[tid+thrd];
odd=thrd%2;
inx=thrd/2;
@@ -2678,16 +2705,21 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
decode_space(space,thread,src1,mem,addr);
type_info_key::type_decode(type,size,t);
- printf("mma_st: thrd=%d,addr=%d, fp16(size=%d), stride=%d\n",thrd,addr_reg.u32,size,src2_data.u32);
- mem->write(addr+inx*2*src2_data.u32+odd*16,size/8,&v[0].s64,thread,pI);
- mem->write(addr+inx*2*src2_data.u32+odd*16+size/8,size/8,&v[1].s64,thread,pI);
- mem->write(addr+inx*2*src2_data.u32+odd*16+2*size/8,size/8,&v[2].s64,thread,pI);
- mem->write(addr+inx*2*src2_data.u32+odd*16+3*size/8,size/8,&v[3].s64,thread,pI);
- mem->write(addr+inx*2*src2_data.u32+odd*16+4*size/8,size/8,&v[4].s64,thread,pI);
- mem->write(addr+inx*2*src2_data.u32+odd*16+5*size/8,size/8,&v[5].s64,thread,pI);
- mem->write(addr+inx*2*src2_data.u32+odd*16+6*size/8,size/8,&v[6].s64,thread,pI);
- mem->write(addr+inx*2*src2_data.u32+odd*16+7*size/8,size/8,&v[7].s64,thread,pI);
-
+ printf("mma_st: thrd=%d,addr=%d, fp(size=%d), stride=%d\n",thrd,addr_reg.u32,size,src2_data.u32);
+ if(type==F16_TYPE){
+ for(k=0;k<8;k++){
+ mem->write(addr+inx*2*src2_data.u32+odd*16+k*size/8,size/8,&v[k].s64,thread,pI);
+ }
+ }
+ else if(type==F32_TYPE){
+ for(k=0;k<8;k++){
+ mem->write(addr+inx*4*src2_data.u32+odd*32+k*size/8,size/8,&v[k].s64,thread,pI);
+ }
+
+ }
+ else{
+ printf("wmma:wrong error type\n");
+ }
printf("wmma:store:thread%d=%x,%x,%x,%x,%x,%x,%x,%x\n",thrd,v[0].s64,v[1].s64,v[2].s64,v[3].s64,v[4].s64,v[5].s64,v[6].s64,v[7].s64);
delete [] v;
@@ -2712,8 +2744,8 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
thread = core->get_thread_info()[tid+thrd];
odd=thrd%2;
inx=thrd/2;
- ptx_reg_t src1_data = thread->get_operand_value(src1, dst, type, thread, 1);
- ptx_reg_t src2_data = thread->get_operand_value(src2, dst, type, thread, 1);
+ ptx_reg_t src1_data = thread->get_operand_value(src1, dst, U32_TYPE, thread, 1);
+ ptx_reg_t src2_data = thread->get_operand_value(src2, dst, U32_TYPE, thread, 1);
memory_space_t space = pI->get_space();
@@ -2726,16 +2758,32 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
ptx_reg_t data1, data2, data3, data4;
ptx_reg_t data5, data6, data7, data8;
printf("mma_ld: thrd=%d,addr=%d, fp16(size=%d), stride=%d\n",thrd,src1_data.u32,size,src2_data.u32);
- mem->read(addr+inx*2*src2_data.u32+odd*16,size/8,&data1.s64);
- mem->read(addr+inx*2*src2_data.u32+odd*16+size/8,size/8,&data2.s64);
- mem->read(addr+inx*2*src2_data.u32+odd*16+2*size/8,size/8,&data3.s64);
- mem->read(addr+inx*2*src2_data.u32+odd*16+3*size/8,size/8,&data4.s64);
- mem->read(addr+inx*2*src2_data.u32+odd*16+4*size/8,size/8,&data5.s64);
- mem->read(addr+inx*2*src2_data.u32+odd*16+5*size/8,size/8,&data6.s64);
- mem->read(addr+inx*2*src2_data.u32+odd*16+6*size/8,size/8,&data7.s64);
- mem->read(addr+inx*2*src2_data.u32+odd*16+7*size/8,size/8,&data8.s64);
+ if(type==F16_TYPE){
+ mem->read(addr+inx*2*src2_data.u32+odd*16,size/8,&data1.s64);
+ mem->read(addr+inx*2*src2_data.u32+odd*16+size/8,size/8,&data2.s64);
+ mem->read(addr+inx*2*src2_data.u32+odd*16+2*size/8,size/8,&data3.s64);
+ mem->read(addr+inx*2*src2_data.u32+odd*16+3*size/8,size/8,&data4.s64);
+ mem->read(addr+inx*2*src2_data.u32+odd*16+4*size/8,size/8,&data5.s64);
+ mem->read(addr+inx*2*src2_data.u32+odd*16+5*size/8,size/8,&data6.s64);
+ mem->read(addr+inx*2*src2_data.u32+odd*16+6*size/8,size/8,&data7.s64);
+ mem->read(addr+inx*2*src2_data.u32+odd*16+7*size/8,size/8,&data8.s64);
+ printf("thread%d=%x,%x,%x,%x,%x,%x,%x,%x\n",0,data1.s64,data2.s64,data3.s64,data4.s64,data5.s64,data6.s64,data7.s64,data8.s64);
+ }
+ else if(type==F32_TYPE){
+ mem->read(addr+inx*4*src2_data.u32+odd*32,size/8,&data1.s64);
+ mem->read(addr+inx*4*src2_data.u32+odd*32+size/8,size/8,&data2.s64);
+ mem->read(addr+inx*4*src2_data.u32+odd*32+2*size/8,size/8,&data3.s64);
+ mem->read(addr+inx*4*src2_data.u32+odd*32+3*size/8,size/8,&data4.s64);
+ mem->read(addr+inx*4*src2_data.u32+odd*32+4*size/8,size/8,&data5.s64);
+ mem->read(addr+inx*4*src2_data.u32+odd*32+5*size/8,size/8,&data6.s64);
+ mem->read(addr+inx*4*src2_data.u32+odd*32+6*size/8,size/8,&data7.s64);
+ mem->read(addr+inx*4*src2_data.u32+odd*32+7*size/8,size/8,&data8.s64);
+ printf("thread%d=%f,%f,%f,%f,%f,%f,%f,%f\n",thrd,data1.f32,data2.f32,data3.f32,data4.f32,data5.f32,data6.f32,data7.f32,data8.f32);
+ }
+ else{
+ printf("wmma_ld:wrong type\n");
+ }
thread->set_wmma_vector_operand_values(dst,data1,data2,data3,data4,data5,data6,data7,data8);
- printf("thread%d=%x,%x,%x,%x,%x,%x,%x,%x\n",0,data1.s64,data2.s64,data3.s64,data4.s64,data5.s64,data6.s64,data7.s64,data8.s64);
thread->m_last_effective_address = addr;
thread->m_last_memory_space = space;
generated by cgit v1.3 (git 2.53.0) at 2026-07-11 04:01:49 +0000