From 4d1d5b569f55dd613bea26714eb1ad931a10be35 Mon Sep 17 00:00:00 2001 From: Max Ryabinin Date: Fri, 1 Jul 2022 17:40:02 +0300 Subject: Reduce diff --- csrc/pythonInterface.c | 115 ++++++++++++++++++++++++------------------------- 1 file changed, 57 insertions(+), 58 deletions(-) (limited to 'csrc/pythonInterface.c') diff --git a/csrc/pythonInterface.c b/csrc/pythonInterface.c index 229b7ed..1f690c5 100644 --- a/csrc/pythonInterface.c +++ b/csrc/pythonInterface.c @@ -38,15 +38,15 @@ MAKE_FUNC32(adagrad, ADAGRAD, half, 16) #define MAKE_FUNC8(fname, oname, gtype, gbits) \ void fname##_static_8bit_g##gbits(gtype* p, gtype* g, unsigned char* state1, unsigned char* state2, \ - float *unorm, float max_unorm, float param_norm, \ + float *unorm, float max_unorm, float param_norm, \ float beta1, float beta2, \ float eps, int step, float lr, \ float* quantiles1, float* quantiles2, \ float* max1, float* max2, float* new_max1, float* new_max2, \ float weight_decay, float gnorm_scale, int n) \ { \ - optimizerStatic8bit(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr, \ - quantiles1, quantiles2, max1, max2, new_max1, new_max2, weight_decay, gnorm_scale, n); \ + optimizerStatic8bit(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr, \ + quantiles1, quantiles2, max1, max2, new_max1, new_max2, weight_decay, gnorm_scale, n); \ } \ MAKE_FUNC8(adam, ADAM, float, 32) @@ -86,37 +86,37 @@ void dequantizeBlockwise_fp32(float *code, unsigned char *A, float *absmax, floa extern "C" { -#if BUILD_CUDA -void cestimate_quantiles_fp32(float *A, float *code, float offset, int n){ estimateQuantiles_fp32(A, code, offset, n); } -void cestimate_quantiles_fp16(half *A, float *code, float offset, int n){ estimateQuantiles_fp16(A, code, offset, n); } -void cquantize(float *code, float *A, unsigned char *out, int n){ quantize(code, A, out, n); } -void cdequantize(float *code, unsigned char *A, float *out, int n){ dequantize(code, A, out, n); } -void cquantize_blockwise_fp16(float * code, half *A, float *absmax, unsigned char *out, const int n){ quantizeBlockwise_fp16(code, A, absmax, out, n); } -void cquantize_blockwise_fp32(float * code, float *A, float *absmax, unsigned char *out, const int n){ quantizeBlockwise_fp32(code, A, absmax, out, n); } -void cquantize_blockwise_stochastic_fp16(float * code, half *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n){ quantizeBlockwise_stochastic_fp16(code, A, absmax, out, rand, rand_offset, n); } -void cquantize_blockwise_stochastic_fp32(float * code, float *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n){ quantizeBlockwise_stochastic_fp32(code, A, absmax, out, rand, rand_offset, n); } - -void cdequantize_blockwise_fp16(float *code, unsigned char *A, float *absmax, half *out, int blocksize, const int n){ dequantizeBlockwise_fp16(code, A, absmax, out, blocksize, n); } -void cdequantize_blockwise_fp32(float *code, unsigned char *A, float *absmax, float *out, int blocksize, const int n){ dequantizeBlockwise_fp32(code, A, absmax, out, blocksize, n); } - -#define MAKE_CFUNC32(name, gtype, gbits) \ - void c##name##32bit_g##gbits(gtype *g, gtype *p, \ - float* state1, float* state2, float *unorm, float max_unorm, float param_norm, \ - const float beta1, const float beta2, const float eps, const float weight_decay, \ - const int step, const float lr, const float gnorm_scale, bool skip_zeros, const int n) \ - { name##32bit_g##gbits(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n); } \ - -MAKE_CFUNC32(adam, float, 32) -MAKE_CFUNC32(adam, half, 16) -MAKE_CFUNC32(momentum, float, 32) -MAKE_CFUNC32(momentum, half, 16) -MAKE_CFUNC32(rmsprop, float, 32) -MAKE_CFUNC32(rmsprop, half, 16) -MAKE_CFUNC32(adagrad, float, 32) -MAKE_CFUNC32(adagrad, half, 16) - -#define MAKE_CFUNC8(name, gtype, gbits) \ - void c##name##_static_8bit_g##gbits(gtype* p, gtype* g, unsigned char* state1, unsigned char* state2, \ + if #BUILD_CUDA + void cestimate_quantiles_fp32(float *A, float *code, float offset, int n){ estimateQuantiles_fp32(A, code, offset, n); } + void cestimate_quantiles_fp16(half *A, float *code, float offset, int n){ estimateQuantiles_fp16(A, code, offset, n); } + void cquantize(float *code, float *A, unsigned char *out, int n){ quantize(code, A, out, n); } + void cdequantize(float *code, unsigned char *A, float *out, int n){ dequantize(code, A, out, n); } + void cquantize_blockwise_fp16(float * code, half *A, float *absmax, unsigned char *out, const int n){ quantizeBlockwise_fp16(code, A, absmax, out, n); } + void cquantize_blockwise_fp32(float * code, float *A, float *absmax, unsigned char *out, const int n){ quantizeBlockwise_fp32(code, A, absmax, out, n); } + void cquantize_blockwise_stochastic_fp16(float * code, half *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n){ quantizeBlockwise_stochastic_fp16(code, A, absmax, out, rand, rand_offset, n); } + void cquantize_blockwise_stochastic_fp32(float * code, float *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n){ quantizeBlockwise_stochastic_fp32(code, A, absmax, out, rand, rand_offset, n); } + + void cdequantize_blockwise_fp16(float *code, unsigned char *A, float *absmax, half *out, int blocksize, const int n){ dequantizeBlockwise_fp16(code, A, absmax, out, blocksize, n); } + void cdequantize_blockwise_fp32(float *code, unsigned char *A, float *absmax, float *out, int blocksize, const int n){ dequantizeBlockwise_fp32(code, A, absmax, out, blocksize, n); } + + #define MAKE_CFUNC32(name, gtype, gbits) \ + void c##name##32bit_g##gbits(gtype *g, gtype *p, \ + float* state1, float* state2, float *unorm, float max_unorm, float param_norm, \ + const float beta1, const float beta2, const float eps, const float weight_decay, \ + const int step, const float lr, const float gnorm_scale, bool skip_zeros, const int n) \ + { name##32bit_g##gbits(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n); } \ + + MAKE_CFUNC32(adam, float, 32) + MAKE_CFUNC32(adam, half, 16) + MAKE_CFUNC32(momentum, float, 32) + MAKE_CFUNC32(momentum, half, 16) + MAKE_CFUNC32(rmsprop, float, 32) + MAKE_CFUNC32(rmsprop, half, 16) + MAKE_CFUNC32(adagrad, float, 32) + MAKE_CFUNC32(adagrad, half, 16) + + #define MAKE_CFUNC8(name, gtype, gbits) \ + void c##name##_static_8bit_g##gbits(gtype* p, gtype* g, unsigned char* state1, unsigned char* state2, \ float *unorm, float max_unorm, float param_norm, \ float beta1, float beta2, \ float eps, int step, float lr, \ @@ -124,40 +124,39 @@ MAKE_CFUNC32(adagrad, half, 16) float* max1, float* max2, float* new_max1, float* new_max2, \ float weight_decay, float gnorm_scale, int n) \ { \ - name##_static_8bit_g##gbits(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr, \ - quantiles1, quantiles2, max1, max2, new_max1, new_max2, weight_decay, gnorm_scale, n); \ + name##_static_8bit_g##gbits(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr, \ + quantiles1, quantiles2, max1, max2, new_max1, new_max2, weight_decay, gnorm_scale, n); \ } \ -MAKE_CFUNC8(adam, float, 32) -MAKE_CFUNC8(adam, half, 16) -MAKE_CFUNC8(momentum, float, 32) -MAKE_CFUNC8(momentum, half, 16) -MAKE_CFUNC8(rmsprop, float, 32) -MAKE_CFUNC8(rmsprop, half, 16) + MAKE_CFUNC8(adam, float, 32) + MAKE_CFUNC8(adam, half, 16) + MAKE_CFUNC8(momentum, float, 32) + MAKE_CFUNC8(momentum, half, 16) + MAKE_CFUNC8(rmsprop, float, 32) + MAKE_CFUNC8(rmsprop, half, 16) -#define MAKE_CBLOCKWISE8(fname, optim_name, gtype, gbits) \ + #define MAKE_CBLOCKWISE8(fname, optim_name, gtype, gbits) \ void c##fname##_8bit_blockwise_fp##gbits(gtype* p, gtype* g, \ unsigned char* state1, unsigned char* state2, float beta1, float beta2, float eps, int step, float lr, \ float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, bool skip_zeros, int n) \ { fname##_8bit_blockwise_fp##gbits(p, g, state1, state2, beta1, beta2, eps, step, lr, quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, skip_zeros, n); } \ -MAKE_CBLOCKWISE8(adam, ADAM, half, 16) -MAKE_CBLOCKWISE8(adam, ADAM, float, 32) -MAKE_CBLOCKWISE8(momentum, MOMENTUM, half, 16) -MAKE_CBLOCKWISE8(momentum, MOMENTUM, float, 32) -MAKE_CBLOCKWISE8(rmsprop, RMSPROP, half, 16) -MAKE_CBLOCKWISE8(rmsprop, RMSPROP, float, 32) -MAKE_CBLOCKWISE8(adagrad, ADAGRAD, half, 16) -MAKE_CBLOCKWISE8(adagrad, ADAGRAD, float, 32) + MAKE_CBLOCKWISE8(adam, ADAM, half, 16) + MAKE_CBLOCKWISE8(adam, ADAM, float, 32) + MAKE_CBLOCKWISE8(momentum, MOMENTUM, half, 16) + MAKE_CBLOCKWISE8(momentum, MOMENTUM, float, 32) + MAKE_CBLOCKWISE8(rmsprop, RMSPROP, half, 16) + MAKE_CBLOCKWISE8(rmsprop, RMSPROP, float, 32) + MAKE_CBLOCKWISE8(adagrad, ADAGRAD, half, 16) + MAKE_CBLOCKWISE8(adagrad, ADAGRAD, float, 32) -void cpercentile_clipping_g32(float * g, float *gnorm_vec, int step, const int n){ percentileClipping_g32(g, gnorm_vec, step, n); } -void cpercentile_clipping_g16(half * g, float *gnorm_vec, int step, const int n){ percentileClipping_g16(g, gnorm_vec, step, n); } -void chistogram_scatter_add_2d(float* histogram, int *index1, int *index2, float *src, int maxidx1, int n){ histogramScatterAdd2D(histogram, index1, index2, src, maxidx1, n); } -#endif + void cpercentile_clipping_g32(float * g, float *gnorm_vec, int step, const int n){ percentileClipping_g32(g, gnorm_vec, step, n); } + void cpercentile_clipping_g16(half * g, float *gnorm_vec, int step, const int n){ percentileClipping_g16(g, gnorm_vec, step, n); } + void chistogram_scatter_add_2d(float* histogram, int *index1, int *index2, float *src, int maxidx1, int n){ histogramScatterAdd2D(histogram, index1, index2, src, maxidx1, n); } -void cquantize_blockwise_cpu_fp32(float *code, float *A, float *absmax, unsigned char *out, const int n){ quantize_cpu(code, A, absmax, out, n); } -void cdequantize_blockwise_cpu_fp32(float *code, unsigned char *A, float *absmax, float *out, const int n){ dequantize_cpu(code, A, absmax, out, n); } + #endif + void cquantize_blockwise_cpu_fp32(float *code, float *A, float *absmax, unsigned char *out, const int n){ quantize_cpu(code, A, absmax, out, n); } + void cdequantize_blockwise_cpu_fp32(float *code, unsigned char *A, float *absmax, float *out, const int n){ dequantize_cpu(code, A, absmax, out, n); } } - -- cgit v1.2.3