summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--cuda-kernels/TensorCoreMatrixCArrangement.xlsxbin0 -> 76737 bytes
-rw-r--r--cuda-kernels/tensorcore_type16_32.cu226
-rw-r--r--src/cuda-sim/cuda-math.h4
-rw-r--r--src/cuda-sim/instructions.cc32
-rw-r--r--src/cuda-sim/ptx_sim.h11
5 files changed, 255 insertions, 18 deletions
diff --git a/cuda-kernels/TensorCoreMatrixCArrangement.xlsx b/cuda-kernels/TensorCoreMatrixCArrangement.xlsx
new file mode 100644
index 0000000..785ecc3
--- /dev/null
+++ b/cuda-kernels/TensorCoreMatrixCArrangement.xlsx
Binary files differ
diff --git a/cuda-kernels/tensorcore_type16_32.cu b/cuda-kernels/tensorcore_type16_32.cu
new file mode 100644
index 0000000..7084bf9
--- /dev/null
+++ b/cuda-kernels/tensorcore_type16_32.cu
@@ -0,0 +1,226 @@
+#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,half *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;
+ wmma::fragment<wmma::accumulator, WMMA_M, WMMA_N, WMMA_K, half> d_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);
+
+// for(int i=0; i < c_frag.num_elements; i++) {
+//// c_frag.x[i]=c_frag.x[i]+c_frag.x[i];
+// float temp=c_frag.x[i];
+// printf("THREAD%d:%d: %f \n",threadIdx.x,i,temp );
+// }
+ wmma::mma_sync(c_frag, a_frag, b_frag, c_frag);
+
+ for(int i=0; i < c_frag.num_elements; i++) {
+ d_frag.x[i]=c_frag.x[i];
+ }
+ wmma::store_matrix_sync(d_fp16, d_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);
+ 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);
+
+ 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_fp16 , MATRIX_M, MATRIX_N, MATRIX_K);
+ cudaErrCheck(cudaEventRecord(stopWMMA));
+ cudaErrCheck(cudaEventSynchronize(stopWMMA));
+
+ convertFp16ToFp32 <<< (MATRIX_M * MATRIX_N + 255) / 256, 256 >>> (d_fp32, d_fp16, MATRIX_K * MATRIX_N);
+ // 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));
+ int t=600000;
+ while(t-->0);
+
+ 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");
+ }
+ int suc=1;
+ for(int m=0;m<MATRIX_M;m++){
+ for(int n=0;n<MATRIX_N;n++){
+ if(abs(d_cal_host_wmma[m*MATRIX_N+n]-d_host_wmma[m*MATRIX_N+n])>1)
+ {
+ printf("ERROR:\n");
+ suc=0;
+ }
+ }
+ }
+ if(suc==1)
+ printf("COMPLETED_SUCCESSFULLY\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/cuda-math.h b/src/cuda-sim/cuda-math.h
index e38e499..314f813 100644
--- a/src/cuda-sim/cuda-math.h
+++ b/src/cuda-sim/cuda-math.h
@@ -322,7 +322,7 @@ float __internal_accurate_fdividef(float a, float b)
float __saturatef(float a)
{
float b;
- if (isnan(a)) b = 0.0f;
+ if (std::isnan(a)) b = 0.0f;
else if (a >= 1.0f) b = 1.0f;
else if (a <= 0.0f) b = 0.0f;
else b = a;
@@ -358,7 +358,7 @@ int __signbitd(double d)
#ifdef __APPLE__
int isnanf(float a)
{
- return (isnan(a));
+ return (std::isnan(a));
}
#endif
diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc
index f314e62..59415f4 100644
--- a/src/cuda-sim/instructions.cc
+++ b/src/cuda-sim/instructions.cc
@@ -1644,9 +1644,9 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
for(k=0;k<8;k++){
if(k%2==0)
- hex_val=((v[k/2].s64&0xffff0000)>>16);
- else
hex_val=(v[k/2].s64&0xffff);
+ else
+ hex_val=((v[k/2].s64&0xffff0000)>>16);
nw_v[k].f16 =*((half *)&hex_val);
matrix_c[row+k][offset]=nw_v[k];
}
@@ -1767,10 +1767,10 @@ void mma_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
}
else if(type==F16_TYPE){
ptx_reg_t nw_data1, nw_data2, nw_data3, nw_data4;
- nw_data1.s64=((matrix_d[row][offset].s64 & 0xffff)<<16)|((matrix_d[row+1][offset].s64&0xffff));
- nw_data2.s64=((matrix_d[row+2][offset].s64 & 0xffff)<<16)|((matrix_d[row+3][offset].s64&0xffff));
- nw_data3.s64=((matrix_d[row+4][offset].s64 & 0xffff)<<16)|((matrix_d[row+5][offset].s64&0xffff));
- nw_data4.s64=((matrix_d[row+6][offset].s64 & 0xffff)<<16)|((matrix_d[row+7][offset].s64&0xffff));
+ nw_data1.s64=((matrix_d[row][offset].s64 & 0xffff))|((matrix_d[row+1][offset].s64&0xffff)<<16);
+ nw_data2.s64=((matrix_d[row+2][offset].s64 & 0xffff))|((matrix_d[row+3][offset].s64&0xffff)<<16);
+ nw_data3.s64=((matrix_d[row+4][offset].s64 & 0xffff))|((matrix_d[row+5][offset].s64&0xffff)<<16);
+ nw_data4.s64=((matrix_d[row+6][offset].s64 & 0xffff))|((matrix_d[row+7][offset].s64&0xffff)<<16);
thread->set_vector_operand_values(dst,nw_data1,nw_data2,nw_data3,nw_data4);
printf("thread%d=%x,%x,%x,%x",thrd,nw_data1.s64,nw_data2.s64,nw_data3.s64,nw_data4.s64);
@@ -2295,7 +2295,7 @@ ptx_reg_t d2d( ptx_reg_t x, unsigned from_width, unsigned to_width, int to_sign,
y.f64 = x.f64;
break;
}
- if (isnan(y.f64)) {
+ if (std::isnan(y.f64)) {
y.u64 = 0xfff8000000000000ull;
} else if (saturation_mode) {
y.f64 = cuda_math::__saturatef(y.f64);
@@ -2426,7 +2426,7 @@ void ptx_round(ptx_reg_t& data, int rounding_mode, int type)
}
}
if ((type == F64_TYPE)||(type == FF64_TYPE)) {
- if (isnan(data.f64)) {
+ if (std::isnan(data.f64)) {
data.u64 = 0xfff8000000000000ull;
}
}
@@ -2828,9 +2828,9 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
ptx_reg_t nw_v[8];
for(k=0;k<8;k++){
if(k%2==0)
- nw_v[k].s64=((v[k/2].s64&0xffff0000)>>16);
- else
nw_v[k].s64=(v[k/2].s64&0xffff);
+ else
+ nw_v[k].s64=((v[k/2].s64&0xffff0000)>>16);
}
for(k=0;k<8;k++){
@@ -2905,10 +2905,10 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
}
else{
ptx_reg_t nw_data1, nw_data2, nw_data3, nw_data4;
- nw_data1.s64=((data1.s64 & 0xffff)<<16)|((data2.s64&0xffff));
- nw_data2.s64=((data3.s64 & 0xffff)<<16)|((data4.s64&0xffff));
- nw_data3.s64=((data5.s64 & 0xffff)<<16)|((data6.s64&0xffff));
- nw_data4.s64=((data7.s64 & 0xffff)<<16)|((data8.s64&0xffff));
+ nw_data1.s64=((data1.s64 & 0xffff))|((data2.s64&0xffff)<<16);
+ nw_data2.s64=((data3.s64 & 0xffff))|((data4.s64&0xffff)<<16);
+ nw_data3.s64=((data5.s64 & 0xffff))|((data6.s64&0xffff)<<16);
+ nw_data4.s64=((data7.s64 & 0xffff))|((data8.s64&0xffff)<<16);
printf("wmma_load:data1.s64=%x,data2.s64=%x,new_data1.s64=%x\n",data1.s64,data2.s64,nw_data1.s64);
printf("wmma_load:data3.s64=%x,data4.s64=%x,new_data2.s64=%x\n",data3.s64,data4.s64,nw_data2.s64);
printf("wmma_load:data5.s64=%x,data6.s64=%x,new_data3.s64=%x\n",data5.s64,data6.s64,nw_data3.s64);
@@ -3139,12 +3139,12 @@ void mad_def( const ptx_instruction *pI, ptx_thread_info *thread, bool use_carry
bool isNaN(float x)
{
- return isnan(x);
+ return std::isnan(x);
}
bool isNaN(double x)
{
- return isnan(x);
+ return std::isnan(x);
}
void max_impl( const ptx_instruction *pI, ptx_thread_info *thread )
diff --git a/src/cuda-sim/ptx_sim.h b/src/cuda-sim/ptx_sim.h
index fbd7881..ea87e59 100644
--- a/src/cuda-sim/ptx_sim.h
+++ b/src/cuda-sim/ptx_sim.h
@@ -42,6 +42,12 @@
#include "memory.h"
+#define GCC_VERSION (__GNUC__ * 10000 \
+ + __GNUC_MINOR__ * 100 \
+ + __GNUC_PATCHLEVEL__)
+
+
+
struct param_t {
const void *pdata;
int type;
@@ -128,7 +134,12 @@ union ptx_reg_t {
unsigned short u16;
unsigned int u32;
unsigned long long u64;
+ //gcc 4.7.0
+ #if GCC_VERSION >= 40700
half f16;
+ #else
+ float f16;
+ #endif
float f32;
double f64;
struct {
generated by cgit v1.3 (git 2.53.0) at 2026-07-11 02:35:09 +0000