diff --git a/Makefile b/Makefile index fe2f26ecb..c7ccf462d 100644 --- a/Makefile +++ b/Makefile @@ -171,15 +171,15 @@ embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o libllama.so: llama.o ggml.o $(CXX) $(CXXFLAGS) -shared -fPIC -o libllama.so llama.o ggml.o $(LDFLAGS) - + # # Tests # benchmark: ggml.o - $(CXX) $(CXXFLAGS) examples/benchmark/benchmark-q4_0-matmult.c ggml.o -o benchmark-q4_0-matmult $(LDFLAGS) + $(CXX) $(CXXFLAGS) examples/benchmark/benchmark-q4_0-matmult.c ggml.o -o benchmark-q4_0-matmult $(LDFLAGS) ./benchmark-q4_0-matmult - + .PHONY: tests tests: bash ./tests/run-tests.sh diff --git a/examples/benchmark/benchmark-q4_0-matmult.c b/examples/benchmark/benchmark-q4_0-matmult.c index 9ca9b133a..90f537fd8 100644 --- a/examples/benchmark/benchmark-q4_0-matmult.c +++ b/examples/benchmark/benchmark-q4_0-matmult.c @@ -24,12 +24,12 @@ float tensor_sum_elements(struct ggml_tensor * tensor) { float sum = 0; - if (tensor->type==6) { - for (int j = 0; j < tensor->ne[1]; j++) { - for (int k = 0; k < tensor->ne[0]; k++) { - sum += ((float *) tensor->data)[j*tensor->ne[0]+k]; - } - } + if (tensor->type==6) { + for (int j = 0; j < tensor->ne[1]; j++) { + for (int k = 0; k < tensor->ne[0]; k++) { + sum += ((float *) tensor->data)[j*tensor->ne[0]+k]; + } + } } return sum; } @@ -39,7 +39,7 @@ float tensor_sum_elements(struct ggml_tensor * tensor) { These are mapping to unknown GGML_TYPE_I8, GGML_TYPE_I16, - GGML_TYPE_I32, + GGML_TYPE_I32, GGML_TYPE_COUNT, */ @@ -50,7 +50,7 @@ float tensor_sum_elements(struct ggml_tensor * tensor) { TENSOR->ne[0], TENSOR->ne[1], TENSOR->ne[2], TENSOR->nb[0], TENSOR->nb[1], TENSOR->nb[2]); \ { float sum = tensor_sum_elements(TENSOR); printf("Sum of tensor %s is %6.2f\n",#TENSOR, sum); } -struct benchmark_params_struct { +struct benchmark_params_struct { int32_t n_threads = 1; int32_t n_iterations = 10; }; @@ -67,7 +67,7 @@ void print_usage(int /*argc*/, char ** argv, struct benchmark_params_struct para int main(int argc, char ** argv) { - + struct benchmark_params_struct benchmark_params; bool invalid_param = false; @@ -90,7 +90,7 @@ int main(int argc, char ** argv) { } else if (arg == "-h" || arg == "--help") { print_usage(argc, argv, benchmark_params); exit(0); - } + } if (invalid_param) { fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str()); print_usage(argc, argv, benchmark_params); @@ -101,9 +101,9 @@ int main(int argc, char ** argv) { // create the ggml context printf("Starting Test\n"); - - + + struct ggml_context * ctx; //const int sizex = 4096; //const int sizey = 11008; @@ -111,31 +111,31 @@ int main(int argc, char ** argv) { #undef VERBOSE_DEBUGGING #ifndef VERBOSE_DEBUGGING const int sizey = 4096; - const int sizex = 11008; + const int sizex = 11008; const int sizez = 128; #else /* Working - let's increase size */ const int sizey = 1; - const int sizex = (8*32); + const int sizex = (8*32); const int sizez = 1; /*const int sizey = 1; - const int sizex = 3*(8*32); + const int sizex = 3*(8*32); const int sizez = 1;*/ #endif //printf("Memsize required = %i\n", sizex*sizex); - ggml_type wtype = GGML_TYPE_F32; - + ggml_type wtype = GGML_TYPE_F32; + size_t ctx_size = 0; ctx_size += sizex*sizey*ggml_type_sizef(wtype); ctx_size += sizex*sizey*ggml_type_sizef(wtype); ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); ctx_size += sizex*sizeof(float); - ctx_size += 1024*1024*100; - + ctx_size += 1024*1024*100; + printf("Allocating Memory of size %li byes, %li MB\n",ctx_size, (ctx_size/1024/1024)); - + struct ggml_init_params params = { /*.mem_size =*/ ctx_size, /*.mem_buffer =*/ NULL, @@ -147,88 +147,88 @@ int main(int argc, char ** argv) { fprintf(stderr, "%s: ggml_init() failed\n", __func__); return false; } - - + + printf("Creating new tensors\n"); // printf("Creating new tensor m1\n"); struct ggml_tensor * m11 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey); ggml_set_f32(m11, 1.0f); - + // printf("Creating new tensor m1\n"); struct ggml_tensor * m12 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey); ggml_set_f32(m12, 1.5f); - + // printf("Creating new tensor m2\n"); struct ggml_tensor * m2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizez); ggml_set_f32(m2, 2.0f); - + printf("\n------ Test 1 - Matrix Mult via F32 code ------------------------------------------------------------------------------\n"); // printf("Creating new tensor m11xm2\n"); struct ggml_tensor * m11xm2 = ggml_mul_mat(ctx, m11, m2); - + // printf("Creating compute graph\n"); struct ggml_cgraph gf = ggml_build_forward(m11xm2); - + gf.n_threads=benchmark_params.n_threads; - printf("cgraph->n_threads=%i\n",gf.n_threads); - + printf("cgraph->n_threads=%i\n",gf.n_threads); + TENSOR_DUMP(m11); TENSOR_DUMP(m2); - + ggml_graph_compute(ctx, &gf); TENSOR_DUMP(gf.nodes[0]); - + printf("\n------ Test 2 - Matrix Mult via Q4_0 code ------------------------------------------------------------------------------\n"); - + int32_t nelements = sizex*sizey; int32_t ne[2] = { sizex, sizey }; - - std::vector hist_cur(1 << 4, 0); + + std::vector hist_cur(1 << 4, 0); // Set up a the benchmark matrices // printf("Creating new tensor q11 & Running quantize\n"); struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, sizex, sizey); ggml_quantize_q4_0((const float *) m11->data, q11->data, nelements, ne[0], hist_cur.data()); - + // Set up a the compute graph // printf("Creating new tensor q31\n"); struct ggml_tensor * q31 = ggml_mul_mat(ctx, q11, m2); - + // printf("Creating compute graph\n"); struct ggml_cgraph gf31 = ggml_build_forward(q31); gf31.n_threads=benchmark_params.n_threads; - - // Set up a second graph computation to make sure we override the CPU cache lines + + // Set up a second graph computation to make sure we override the CPU cache lines // printf("Creating new tensor q12 & Running quantize\n"); struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, sizex, sizey); ggml_quantize_q4_0((const float *) m12->data, q12->data, nelements, ne[0], hist_cur.data()); // printf("Creating new tensor q32\n"); struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2); - + //printf("Creating compute graph\n"); struct ggml_cgraph gf32 = ggml_build_forward(q32); gf32.n_threads=benchmark_params.n_threads; - printf("cgraph->n_threads=%i\n",gf31.n_threads); - + printf("cgraph->n_threads=%i\n",gf31.n_threads); + const int dimx = sizex; const int dimy = sizey; const int dimz = sizez; long long int flops_per_dot_product = dimy + dimy; long long int flops_per_matrix = flops_per_dot_product * dimx * dimz; ; printf("Matrix Multiplication of (%i,%i,%i) x (%i,%i,%i) - aboout %6.2f gFLOPS\n\n", sizex, sizey, 1, sizex, sizez, 1, 1.0f*flops_per_matrix / 1000 / 1000 / 1000); - + // Let's use the F32 result from above as a reference for the q4_0 multiplication float sum_of_F32_reference = tensor_sum_elements(gf.nodes[0]); - + printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; FLOPS_per_u_Second\n"); printf("==============================================================================================\n"); - + for (int i=0;i allowed_delta) { printf("\nABORT - ERROR in Matrix Multiplication result - expected %6.2f, got %6.2f (delta %6.2f > allowed_delta %6.2f)\n", - sum_of_F32_reference, + sum_of_F32_reference, sum_of_Q4_result, delta, allowed_delta ); exit(0); } - - // Running a different graph computation to make sure we override the CPU cache lines + + // Running a different graph computation to make sure we override the CPU cache lines ggml_graph_compute(ctx, &gf32); - + } - + }