implemented prmt and started working on variable precision mul inst

7 files changed, 362 insertions, 13 deletions

diff --git a/cuda-kernels/v4p_kernel.cu b/cuda-kernels/v4p_kernel.cu
new file mode 100644
index 0000000..bb9064b
--- /dev/null
+++ b/cuda-kernels/v4p_kernel.cu
@@ -0,0 +1,232 @@
+#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];
+   }
+}
+
+__global__ void convertInt32ToInt8 (int *out, int *in, int n) {
+   int idx = blockDim.x * blockIdx.x + threadIdx.x;
+   if (idx < n/4) {
+      		out[idx] =(in[4*idx]&0xff)|(in[4*idx+1]&0xff)<<8|(in[4*idx+2]&0xff)<<16|(in[4*idx+3]&0xff)<<24;
+   }
+}
+
+__global__ void convertInt8ToInt32 (int *out, int *in, int n) {
+   int idx = blockDim.x * blockIdx.x + threadIdx.x;
+   int shft_amt=8*(idx%4);
+   int shft_mask=0xff<<shft_amt;
+   if (idx < n) {
+      	out[idx]= (in[idx/4]&shft_mask)>>shft_amt;
+   }
+}
+
+int main(int argc, char* argv[]) {
+   int *a_int32;
+   int *b_int32;
+   int *c_int32;
+   int *d_int32;
+
+   int *a_int8;
+   int *b_int8;
+   
+   int  *a_host_wmma;
+   int  *b_host_wmma;
+   int  *c_host_wmma;
+   int  *d_host_wmma;
+   int  *d_cal_host_wmma;
+
+   cudaEvent_t startWMMA;
+   cudaEvent_t stopWMMA;
+   
+   
+   cudaErrCheck(cudaEventCreate(&startWMMA));
+   cudaErrCheck(cudaEventCreate(&stopWMMA));
+   
+   // Use tensor cores
+   cudaErrCheck(cudaMalloc((void**)&a_int32, MATRIX_M * MATRIX_K * sizeof(int)));
+   cudaErrCheck(cudaMalloc((void**)&b_int32, MATRIX_K * MATRIX_N * sizeof(int)));
+   cudaErrCheck(cudaMalloc((void**)&c_int32, MATRIX_K * MATRIX_N * sizeof(int)));
+   cudaErrCheck(cudaMalloc((void**)&d_int32, MATRIX_K * MATRIX_N * sizeof(int)));
+   cudaErrCheck(cudaMalloc((void**)&a_int8, MATRIX_M * MATRIX_K * sizeof(int)/4));
+   cudaErrCheck(cudaMalloc((void**)&b_int8, MATRIX_K * MATRIX_N * sizeof(int)/4));
+
+
+   a_host_wmma      = (int *)malloc(MATRIX_M * MATRIX_K * sizeof(int));
+   b_host_wmma      = (int *)malloc(MATRIX_K * MATRIX_N * sizeof(int));
+   c_host_wmma      = (int *)malloc(MATRIX_M * MATRIX_N * sizeof(int));
+   d_host_wmma      = (int *)malloc(MATRIX_M * MATRIX_N * sizeof(int));
+   d_cal_host_wmma      = (int *)malloc(MATRIX_M * MATRIX_N * sizeof(int));
+
+   printf("a_int32\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)%16;
+		printf("%d ",a_host_wmma[m*MATRIX_K+n]);
+	}
+	printf(";\n");
+   }
+  
+   printf("b_int32\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)%2;
+		printf("%d ",b_host_wmma[m*MATRIX_N+n]);
+	}
+		printf(";\n");
+   }
+   
+   printf("c_int32\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)%2;
+		d_cal_host_wmma[m*MATRIX_N+n]=0;
+		printf("%d ",c_host_wmma[m*MATRIX_N+n]);
+	}
+		printf(";\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_int32,a_host_wmma,  MATRIX_M * MATRIX_K * sizeof(int), cudaMemcpyHostToDevice));
+   cudaErrCheck(cudaMemcpy(b_int32,b_host_wmma,  MATRIX_K * MATRIX_N * sizeof(int), cudaMemcpyHostToDevice));
+   cudaErrCheck(cudaMemcpy(c_int32,c_host_wmma,  MATRIX_M * MATRIX_N * sizeof(int), cudaMemcpyHostToDevice));
+
+   convertInt32ToInt8 <<< (MATRIX_M * MATRIX_K + 255) / 256, 256 >>> (a_int8, a_int32, MATRIX_M * MATRIX_K);
+   convertInt8ToInt32 <<< (MATRIX_M * MATRIX_K + 255) / 256, 256 >>> (d_int32, a_int8, 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);
+   cudaErrCheck(cudaMemcpy(d_host_wmma, d_int32, MATRIX_M * MATRIX_N * sizeof(float), cudaMemcpyDeviceToHost));
+
+
+//AAMIR   printf("\nM = %d, N = %d, K = %d. \n", MATRIX_M, MATRIX_N, MATRIX_K);
+//AAMIR   
+//AAMIR   printf("Running with wmma...\n");
+//AAMIR   cudaErrCheck(cudaEventRecord(startWMMA));
+//AAMIR   wmma_example <<< 1, 32>>> (a_fp16, b_fp16, c_fp32, d_fp32 , MATRIX_M, MATRIX_N, MATRIX_K);
+//AAMIR   cudaErrCheck(cudaEventRecord(stopWMMA));
+//AAMIR   cudaErrCheck(cudaEventSynchronize(stopWMMA));
+//AAMIR
+//AAMIR   // Error checking
+//AAMIR   printf("\nChecking results...\n");
+//AAMIR   cudaErrCheck(cudaMemcpy(d_host_wmma, d_fp32, MATRIX_M * MATRIX_N * sizeof(float), cudaMemcpyDeviceToHost));
+//AAMIR   
+//AAMIR   printf("Results verified: cublas and WMMA agree.\n\n");
+//AAMIR   float wmmaTime;
+//AAMIR   cudaErrCheck(cudaEventElapsedTime(&wmmaTime, startWMMA, stopWMMA));
+//AAMIR   printf("wmma took %fms\n", wmmaTime);
+//AAMIR   
+//AAMIR   cudaErrCheck(cudaEventDestroy(startWMMA));
+//AAMIR   cudaErrCheck(cudaEventDestroy(stopWMMA));
+//AAMIR
+//AAMIR   int t=200000;
+//AAMIR   while(t-->0);
+//AAMIR   printf("D_CALCULATED\n");
+//AAMIR
+//AAMIR   for(int m=0;m<MATRIX_M;m++){
+//AAMIR	for(int n=0;n<MATRIX_N;n++){
+//AAMIR		printf("%.2f,",d_cal_host_wmma[m*MATRIX_N+n]);
+//AAMIR	}
+//AAMIR	printf("\n");
+//AAMIR   }
+   printf("D_WMMA\n");
+   for(int m=0;m<MATRIX_M;m++){
+	for(int n=0;n<MATRIX_N;n++){
+		printf("%d,",d_host_wmma[m*MATRIX_N+n]);
+	}
+	printf("\n");
+    }
+//AAMIR   int suc=1;
+//AAMIR   for(int m=0;m<MATRIX_M;m++){
+//AAMIR	for(int n=0;n<MATRIX_N;n++){
+//AAMIR 		if(abs(d_cal_host_wmma[m*MATRIX_N+n]-d_host_wmma[m*MATRIX_N+n])>1) 
+//AAMIR		{	
+//AAMIR			printf("ERROR:\n");
+//AAMIR			suc=0;
+//AAMIR		}
+//AAMIR	}
+//AAMIR   }
+//AAMIR   if(suc==1)
+//AAMIR	printf("COMPLETED_SUCCESSFULLY\n");
+//AAMIR   
+   
+   cudaErrCheck(cudaFree(a_int32));
+   cudaErrCheck(cudaFree(b_int32));
+   cudaErrCheck(cudaFree(c_int32));
+   cudaErrCheck(cudaFree(d_int32));
+   cudaErrCheck(cudaFree(a_int8));
+   cudaErrCheck(cudaFree(b_int8));
+
+   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 7af157f..aaee2a2 100644
--- a/src/cuda-sim/instructions.cc
+++ b/src/cuda-sim/instructions.cc
@@ -48,7 +48,7 @@
 using half_float::half;
 
 unsigned ptx_instruction::g_num_ptx_inst_uid=0;
-bool g_debug_instruction = 1;
+bool g_debug_instruction = 0;
 
 
 const char *g_opcode_string[NUM_OPCODES] = {
@@ -3911,7 +3911,94 @@ void popc_impl( const ptx_instruction *pI, ptx_thread_info *thread )
 }
 void prefetch_impl( const ptx_instruction *pI, ptx_thread_info *thread ) { inst_not_implemented(pI); }
 void prefetchu_impl( const ptx_instruction *pI, ptx_thread_info *thread ) { inst_not_implemented(pI); }
-void prmt_impl( const ptx_instruction *pI, ptx_thread_info *thread ) { inst_not_implemented(pI); }
+
+int prmt_mode_present(int mode)
+{
+	int returnval=0;
+	switch(mode){
+		case PRMT_F4E_MODE:
+		case PRMT_B4E_MODE:
+		case PRMT_RC8_MODE:
+		case PRMT_RC16_MODE:
+		case PRMT_ECL_MODE:
+		case PRMT_ECR_MODE:	
+			returnval=1;
+			break;
+		default:	
+			break;
+	}
+	return returnval;
+}
+int read_byte(int mode,int control,int d_sel_index,signed long long value){
+
+	int returnval;
+	int prmt_f4e_mode[4][4]={{0,1,2,3},{1,2,3,4},{2,3,4,5},{3,4,5,6}};
+	int prmt_b4e_mode[4][4]={{0,7,6,5},{1,0,7,6},{2,1,0,7},{3,2,1,0}};
+	int prmt_rc8_mode[4][4]={{0,0,0,0},{1,1,1,1},{2,2,2,2},{3,3,3,3}};
+	int prmt_ecl_mode[4][4]={{0,1,2,3},{1,1,2,3},{2,2,2,3},{3,3,3,3}};
+	int prmt_ecr_mode[4][4]={{0,0,0,0},{0,1,1,1},{0,1,2,2},{0,1,2,3}};
+	int prmt_rc16_mode[4][4]={{0,1,0,1},{2,3,2,3},{0,1,0,1},{2,3,2,3}};
+
+	if(!prmt_mode_present(mode)){
+		if(control&0x8){
+			returnval=0xff;
+		}
+		else{
+			returnval= (value>>(8*control)) & 0xff;
+		}
+	}
+	else{
+		switch(mode){
+			case PRMT_F4E_MODE:	returnval=prmt_f4e_mode[control][d_sel_index];break;
+			case PRMT_B4E_MODE:	returnval=prmt_b4e_mode[control][d_sel_index];break;
+			case PRMT_RC8_MODE:	returnval=prmt_rc8_mode[control][d_sel_index];break;
+			case PRMT_ECL_MODE:	returnval=prmt_ecl_mode[control][d_sel_index];break;
+			case PRMT_ECR_MODE:	returnval=prmt_ecr_mode[control][d_sel_index];break;
+			case PRMT_RC16_MODE:	returnval=prmt_rc16_mode[control][d_sel_index];break;
+			default: printf("ERROR\n");break;
+		}
+	}	
+	return (returnval<<8*d_sel_index);
+}
+
+void prmt_impl( const ptx_instruction *pI, ptx_thread_info *thread ) { 
+   
+   ptx_reg_t src1_data, src2_data, src3_data,tmpdata,data;
+   const operand_info &dst  = pI->dst();  
+   const operand_info &src1 = pI->src1();
+   const operand_info &src2 = pI->src2();
+   const operand_info &src3 = pI->src3();
+
+   unsigned mode   = pI->prmt_op();
+   unsigned i_type = pI->get_type();
+
+   src1_data = thread->get_operand_value(src1, dst, i_type, thread, 1);
+   src2_data = thread->get_operand_value(src2, dst, i_type, thread, 1);
+   src3_data = thread->get_operand_value(src3, dst, i_type, thread, 1);
+
+   tmpdata.s64=src1_data.s32|(src2_data.s64<<32);
+   int ctl[4];
+
+   if(!prmt_mode_present(mode)){
+	ctl[0]=(src3_data.s32>>0)&0xf;
+	ctl[1]=(src3_data.s32>>4)&0xf;
+	ctl[2]=(src3_data.s32>>8)&0xf;
+	ctl[3]=(src3_data.s32>>12)&0xf;
+   }
+   else{
+	ctl[0]=ctl[1]=ctl[2]=ctl[3]=(src3_data.s32>>0)&0x3;	
+   }
+   
+   data.s32=0;
+   data.s32=data.s32|read_byte(mode,ctl[0],0,tmpdata.s64);   //First byte-0
+   data.s32=data.s32|read_byte(mode,ctl[1],1,tmpdata.s64);   //Second byte-1
+   data.s32=data.s32|read_byte(mode,ctl[2],2,tmpdata.s64);   //Third byte-2
+   data.s32=data.s32|read_byte(mode,ctl[3],3,tmpdata.s64);   //Fourth byte-3
+	
+   thread->set_operand_value(dst,data, i_type, thread, pI);
+
+
+}
 
 void rcp_impl( const ptx_instruction *pI, ptx_thread_info *thread ) 
 { 
diff --git a/src/cuda-sim/ptx.l b/src/cuda-sim/ptx.l
index 03d6838..a6b6fcc 100644
--- a/src/cuda-sim/ptx.l
+++ b/src/cuda-sim/ptx.l
@@ -164,7 +164,12 @@ breakaddr  TC; ptx_lval.int_value = BREAKADDR_OP; return OPCODE;
 \.row TC; ptx_lval.int_value = ROW; return LAYOUT;
 \.col TC; ptx_lval.int_value = COL; return LAYOUT;
 \.m16n16k16 TC; ptx_lval.int_value = M16N16K16; return CONFIGURATION;
-
+\.f4e TC;  return PRMT_F4E_MODE;
+\.b4e TC;  return PRMT_B4E_MODE;
+\.rc8 TC;  return PRMT_RC8_MODE;
+\.ecl TC;  return PRMT_ECL_MODE;
+\.ecr TC;  return PRMT_ECR_MODE;
+\.rc16 TC; return PRMT_RC16_MODE;
 
 \.align TC; return ALIGN_DIRECTIVE;
 \.branchtargets TC; return BRANCHTARGETS_DIRECTIVE;
diff --git a/src/cuda-sim/ptx.y b/src/cuda-sim/ptx.y
index 8744663..cd455dd 100644
--- a/src/cuda-sim/ptx.y
+++ b/src/cuda-sim/ptx.y
@@ -50,7 +50,6 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 %token  PTR_DIRECTIVE
 %token  ENTRY_DIRECTIVE
 %token  EXTERN_DIRECTIVE
-%token  WEAK_DIRECTIVE
 %token  FILE_DIRECTIVE
 %token  FUNC_DIRECTIVE
 %token  GLOBAL_DIRECTIVE
@@ -203,6 +202,12 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 %token	DOWN_OPTION;
 %token	BFLY_OPTION;
 %token	IDX_OPTION;
+%token	PRMT_F4E_MODE;
+%token	PRMT_B4E_MODE;
+%token	PRMT_RC8_MODE;
+%token	PRMT_RC16_MODE;
+%token	PRMT_ECL_MODE;
+%token	PRMT_ECR_MODE;
 
 %type <int_value> function_decl_header
 %type <ptr_value> function_decl
@@ -432,6 +437,7 @@ option: type_spec
 	| addressable_spec
 	| rounding_mode
 	| wmma_spec 
+	| prmt_spec 
 	| SYNC_OPTION { add_option(SYNC_OPTION); }	
 	| ARRIVE_OPTION { add_option(ARRIVE_OPTION); }
 	| RED_OPTION { add_option(RED_OPTION); }	
@@ -520,6 +526,14 @@ compare_spec:EQ_OPTION { add_option(EQ_OPTION); }
 	| NAN_OPTION { add_option(NAN_OPTION); } 
 	;
 
+prmt_spec: PRMT_F4E_MODE { add_option( PRMT_F4E_MODE); }
+	|  PRMT_B4E_MODE { add_option( PRMT_B4E_MODE); }
+	|  PRMT_RC8_MODE { add_option( PRMT_RC8_MODE); }
+	|  PRMT_RC16_MODE{ add_option( PRMT_RC16_MODE);}
+	|  PRMT_ECL_MODE { add_option( PRMT_ECL_MODE); }
+	|  PRMT_ECR_MODE { add_option( PRMT_ECR_MODE); }
+	;
+
 wmma_spec: WMMA_DIRECTIVE LAYOUT CONFIGURATION{add_space_spec(global_space,0);add_ptr_spec(global_space); add_wmma_option($1);add_wmma_option($2);add_wmma_option($3);}
 	| WMMA_DIRECTIVE LAYOUT LAYOUT CONFIGURATION{add_wmma_option($1);add_wmma_option($2);add_wmma_option($3);add_wmma_option($4);}
 	;
diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc
index 9c2ac69..d12c741 100644
--- a/src/cuda-sim/ptx_ir.cc
+++ b/src/cuda-sim/ptx_ir.cc
@@ -1233,16 +1233,25 @@ ptx_instruction::ptx_instruction( int opcode,
       case HALF_OPTION:
          m_inst_size = 4; // bytes
          break;
-	  case EXTP_OPTION:
-		 break;
-	  case NC_OPTION:
-		 break;
-	  case UP_OPTION:
-	  case DOWN_OPTION:
-	  case BFLY_OPTION:
-	  case IDX_OPTION:
+      case EXTP_OPTION:
+	 break;
+      case NC_OPTION:
+	 break;
+      case UP_OPTION:
+      case DOWN_OPTION:
+      case BFLY_OPTION:
+      case IDX_OPTION:
 		  m_shfl_op = last_ptx_inst_option;
 		  break;
+
+      case PRMT_F4E_MODE:
+      case PRMT_B4E_MODE:
+      case PRMT_RC8_MODE:
+      case PRMT_ECL_MODE:
+      case PRMT_ECR_MODE:
+      case PRMT_RC16_MODE:
+		  m_prmt_op = last_ptx_inst_option;
+		  break;
       default:
          assert(0);
          break;
diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h
index cff312b..cb4556e 100644
--- a/src/cuda-sim/ptx_ir.h
+++ b/src/cuda-sim/ptx_ir.h
@@ -1086,6 +1086,7 @@ public:
    unsigned dimension() const { return m_geom_spec;}
    unsigned barrier_op() const {return m_barrier_op;}
    unsigned shfl_op() const {return m_shfl_op;}
+   unsigned prmt_op() const {return m_prmt_op;}
    enum vote_mode_t { vote_any, vote_all, vote_uni, vote_ballot };
    enum vote_mode_t vote_mode() const { return m_vote_mode; }
 
@@ -1156,6 +1157,7 @@ private:
    unsigned            m_saturation_mode;
    unsigned 		   m_barrier_op;
    unsigned			   m_shfl_op;
+   unsigned			   m_prmt_op;
 
    std::list<int>          m_scalar_type;
    memory_space_t m_space_spec;
diff --git a/src/cuda-sim/ptx_parser.cc b/src/cuda-sim/ptx_parser.cc
index eb81961..6757091 100644
--- a/src/cuda-sim/ptx_parser.cc
+++ b/src/cuda-sim/ptx_parser.cc
@@ -39,7 +39,7 @@ void set_ptx_warp_size(const struct core_config * warp_size)
    g_shader_core_config=warp_size;
 }
 
-static bool g_debug_ir_generation=true;
+static bool g_debug_ir_generation=false;
 const char *g_filename;
 unsigned g_max_regs_per_thread = 0;