diff options
| -rw-r--r-- | csrc/common.h | 2 | ||||
| -rw-r--r-- | csrc/cpu_ops.cpp | 2 | ||||
| -rw-r--r-- | csrc/ops.cu | 290 |
3 files changed, 129 insertions, 165 deletions
diff --git a/csrc/common.h b/csrc/common.h index 35f2463..2f25a58 100644 --- a/csrc/common.h +++ b/csrc/common.h @@ -20,4 +20,4 @@ struct quantize_block_args { void *quantize_block(void *arguments); -#endif
\ No newline at end of file +#endif diff --git a/csrc/cpu_ops.cpp b/csrc/cpu_ops.cpp index 11a2615..89de52d 100644 --- a/csrc/cpu_ops.cpp +++ b/csrc/cpu_ops.cpp @@ -54,4 +54,4 @@ void quantize_cpu(float *code, float *A, float *absmax, unsigned char *out, int for (int i = 0; i < num_blocks; i++) free(args[i]); free(args); -}
\ No newline at end of file +} diff --git a/csrc/ops.cu b/csrc/ops.cu index 464ea2e..b2a1105 100644 --- a/csrc/ops.cu +++ b/csrc/ops.cu @@ -46,103 +46,100 @@ void dequantize(float *code, unsigned char *A, float *out, int n) { CUDA_CHECK_RETURN(cudaPeekAtLastError()); } -template<typename T, int STOCHASTIC> -void quantizeBlockwise(float *code, T *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n) { - int blocks = n / 4096; - blocks = n % 4096 == 0 ? blocks : blocks + 1; - kQuantizeBlockwise < T, 4096, 4, STOCHASTIC ><<<blocks, 1024>>>(code, A, absmax, out, rand, rand_offset, n); - CUDA_CHECK_RETURN(cudaPeekAtLastError()); +template <typename T, int STOCHASTIC> void quantizeBlockwise(float * code, T *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n) +{ + int blocks = n/4096; + blocks = n % 4096 == 0 ? blocks : blocks + 1; + kQuantizeBlockwise<T, 4096, 4, STOCHASTIC><<<blocks, 1024>>>(code, A, absmax, out, rand, rand_offset, n); + CUDA_CHECK_RETURN(cudaPeekAtLastError()); } -template<typename T> -void dequantizeBlockwise(float *code, unsigned char *A, float *absmax, T *out, int blocksize, const int n) { - int blocks = n / blocksize; - blocks = n % blocksize == 0 ? blocks : blocks + 1; - if (blocksize == 4096) - kDequantizeBlockwise < T, 4096, 1024, 4 ><<<blocks, 4096 / 4>>>(code, A, absmax, out, n); - else if (blocksize == 2048) - kDequantizeBlockwise < T, 2048, 512, 4 ><<<blocks, 2048 / 4>>>(code, A, absmax, out, n); - CUDA_CHECK_RETURN(cudaPeekAtLastError()); +template<typename T> void dequantizeBlockwise(float *code, unsigned char *A, float *absmax, T *out, int blocksize, const int n) +{ + int blocks = n/blocksize; + blocks = n % blocksize == 0 ? blocks : blocks + 1; + if(blocksize == 4096) + kDequantizeBlockwise<T, 4096, 1024, 4><<<blocks, 4096/4>>>(code, A, absmax, out, n); + else if(blocksize == 2048) + kDequantizeBlockwise<T, 2048, 512, 4><<<blocks, 2048/4>>>(code, A, absmax, out, n); + CUDA_CHECK_RETURN(cudaPeekAtLastError()); } -template<typename T, int OPTIMIZER> -void optimizer32bit(T *g, T *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) { - int blocks = n / 4096; - blocks = n % 4096 == 0 ? blocks : blocks + 1; - switch (OPTIMIZER) { - case ADAM: - if (max_unorm > 0.0f) { - CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1 * sizeof(float))); - kPreconditionOptimizer32bit2State < T, OPTIMIZER, 4096, - 8 ><<<blocks, 512>>>(g, p, state1, state2, unorm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, n); - CUDA_CHECK_RETURN(cudaPeekAtLastError()); - } - kOptimizer32bit2State < T, - OPTIMIZER ><<<blocks, 1024>>>(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n); - CUDA_CHECK_RETURN(cudaPeekAtLastError()); - break; - case MOMENTUM: - case RMSPROP: - case ADAGRAD: - - if (max_unorm > 0.0f) { - CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1 * sizeof(float))); - kPreconditionOptimizer32bit1State < T, OPTIMIZER, 4096, - 8 ><<<blocks, 512>>>(g, p, state1, unorm, beta1, eps, weight_decay, step, lr, gnorm_scale, n); - CUDA_CHECK_RETURN(cudaPeekAtLastError()); - } - - kOptimizer32bit1State < T, - OPTIMIZER ><<<blocks, 1024>>>(g, p, state1, unorm, max_unorm, param_norm, beta1, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n); - CUDA_CHECK_RETURN(cudaPeekAtLastError()); - break; - } +template<typename T, int OPTIMIZER> void optimizer32bit(T* g, T* 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) +{ + int blocks = n/4096; + blocks = n % 4096 == 0 ? blocks : blocks + 1; + switch(OPTIMIZER) + { + case ADAM: + if(max_unorm > 0.0f) + { + CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1*sizeof(float))); + kPreconditionOptimizer32bit2State<T, OPTIMIZER, 4096, 8><<<blocks, 512>>>(g, p, state1, state2, unorm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, n); + CUDA_CHECK_RETURN(cudaPeekAtLastError()); + } + kOptimizer32bit2State<T, OPTIMIZER><<<blocks, 1024>>>(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n); + CUDA_CHECK_RETURN(cudaPeekAtLastError()); + break; + case MOMENTUM: + case RMSPROP: + case ADAGRAD: + + if(max_unorm > 0.0f) + { + CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1*sizeof(float))); + kPreconditionOptimizer32bit1State<T, OPTIMIZER, 4096, 8><<<blocks, 512>>>(g, p, state1, unorm, beta1, eps, weight_decay, step, lr, gnorm_scale, n); + CUDA_CHECK_RETURN(cudaPeekAtLastError()); + } + + kOptimizer32bit1State<T, OPTIMIZER><<<blocks, 1024>>>(g, p, state1, unorm, max_unorm, param_norm, beta1, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n); + CUDA_CHECK_RETURN(cudaPeekAtLastError()); + break; + } } -template<typename T, int OPTIMIZER> -void optimizerStatic8bit(T *p, T *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, - float *quantiles1, float *quantiles2, - float *max1, float *max2, float *new_max1, float *new_max2, - float weight_decay, - const float gnorm_scale, int n) { - int blocks = n / 4096; - blocks = n % 4096 == 0 ? blocks : blocks + 1; - - if (max_unorm > 0.0f) { CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1 * sizeof(float))); } - - switch (OPTIMIZER) { - case ADAM: - CUDA_CHECK_RETURN(cudaMemset(new_max1, 0, 1 * sizeof(float))); - CUDA_CHECK_RETURN(cudaMemset(new_max2, 0, 1 * sizeof(float))); - kPreconditionOptimizerStatic8bit2State < T, - OPTIMIZER ><<<blocks, 256>>>(p, g, state1, state2, unorm, beta1, beta2, eps, step, quantiles1, quantiles2, max1, max2, new_max1, new_max2, gnorm_scale, n); - CUDA_CHECK_RETURN(cudaPeekAtLastError()); - kOptimizerStatic8bit2State < T, - OPTIMIZER ><<<blocks, 1024>>>(p, g, 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); - CUDA_CHECK_RETURN(cudaPeekAtLastError()); - break; - case MOMENTUM: - case RMSPROP: - case ADAGRAD: - CUDA_CHECK_RETURN(cudaMemset(new_max1, 0, 1 * sizeof(float))); - kPreconditionOptimizerStatic8bit1State < T, - OPTIMIZER ><<<blocks, 256>>>(p, g, state1, unorm, beta1, eps, step, quantiles1, max1, new_max1, weight_decay, gnorm_scale, n); - CUDA_CHECK_RETURN(cudaPeekAtLastError()); - kOptimizerStatic8bit1State < T, OPTIMIZER ><<<blocks, 1024>>>(p, g, state1, unorm, max_unorm, param_norm, beta1, eps, step, lr, - quantiles1, max1, new_max1, weight_decay, gnorm_scale, n); - CUDA_CHECK_RETURN(cudaPeekAtLastError()); - break; - default: - break; - } +template<typename T, int OPTIMIZER> void optimizerStatic8bit(T* p, T* 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, + float* quantiles1, float* quantiles2, + float* max1, float* max2, float* new_max1, float* new_max2, + float weight_decay, + const float gnorm_scale, int n) +{ + int blocks = n/4096; + blocks = n % 4096 == 0 ? blocks : blocks + 1; + + if(max_unorm > 0.0f){ CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1*sizeof(float))); } + + switch(OPTIMIZER) + { + case ADAM: + CUDA_CHECK_RETURN(cudaMemset(new_max1, 0, 1*sizeof(float))); + CUDA_CHECK_RETURN(cudaMemset(new_max2, 0, 1*sizeof(float))); + kPreconditionOptimizerStatic8bit2State<T, OPTIMIZER><<<blocks, 256>>>(p, g, state1, state2, unorm, beta1, beta2, eps, step, quantiles1, quantiles2, max1, max2, new_max1, new_max2, gnorm_scale, n); + CUDA_CHECK_RETURN(cudaPeekAtLastError()); + kOptimizerStatic8bit2State<T, OPTIMIZER><<<blocks, 1024>>>(p, g, 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); + CUDA_CHECK_RETURN(cudaPeekAtLastError()); + break; + case MOMENTUM: + case RMSPROP: + case ADAGRAD: + CUDA_CHECK_RETURN(cudaMemset(new_max1, 0, 1*sizeof(float))); + kPreconditionOptimizerStatic8bit1State<T, OPTIMIZER><<<blocks, 256>>>(p, g, state1, unorm, beta1, eps, step, quantiles1, max1, new_max1, weight_decay, gnorm_scale, n); + CUDA_CHECK_RETURN(cudaPeekAtLastError()); + kOptimizerStatic8bit1State<T, OPTIMIZER><<<blocks, 1024>>>(p, g, state1, unorm, max_unorm, param_norm, beta1, eps, step, lr, + quantiles1, max1, new_max1, weight_decay, gnorm_scale, n); + CUDA_CHECK_RETURN(cudaPeekAtLastError()); + break; + default: + break; + } } #define BLOCKSIZE_2STATE 2048 @@ -150,43 +147,42 @@ void optimizerStatic8bit(T *p, T *g, #define BLOCKSIZE_1STATE 2048 #define NUM_1STATE 8 -template<typename T, int OPTIMIZER> -void optimizerStatic8bitBlockwise(T *p, T *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) { - - int blocks = 0; - switch (OPTIMIZER) { - case ADAM: - blocks = n / BLOCKSIZE_2STATE; - blocks = n % BLOCKSIZE_2STATE == 0 ? blocks : blocks + 1; - kOptimizerStatic8bit2StateBlockwise < T, OPTIMIZER, BLOCKSIZE_2STATE, NUM_2STATE ><<<blocks, BLOCKSIZE_2STATE / - NUM_2STATE>>>(p, g, state1, state2, beta1, beta2, eps, step, lr, - quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, skip_zeros, n); - CUDA_CHECK_RETURN(cudaPeekAtLastError()); - break; - case MOMENTUM: - case RMSPROP: - case ADAGRAD: - blocks = n / BLOCKSIZE_1STATE; - blocks = n % BLOCKSIZE_1STATE == 0 ? blocks : blocks + 1; - kOptimizerStatic8bit1StateBlockwise < T, OPTIMIZER, BLOCKSIZE_1STATE, NUM_1STATE ><<<blocks, BLOCKSIZE_1STATE / - NUM_1STATE>>>(p, g, state1, beta1, beta2, eps, step, lr, - quantiles1, absmax1, weight_decay, gnorm_scale, skip_zeros, n); - CUDA_CHECK_RETURN(cudaPeekAtLastError()); - break; - } +template<typename T, int OPTIMIZER> void optimizerStatic8bitBlockwise(T* p, T* 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) +{ + + int blocks = 0; + switch(OPTIMIZER) + { + case ADAM: + blocks = n/BLOCKSIZE_2STATE; + blocks = n % BLOCKSIZE_2STATE == 0 ? blocks : blocks + 1; + kOptimizerStatic8bit2StateBlockwise<T, OPTIMIZER, BLOCKSIZE_2STATE, NUM_2STATE><<<blocks, BLOCKSIZE_2STATE/NUM_2STATE>>>(p, g, state1, state2, beta1, beta2, eps, step, lr, + quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, skip_zeros, n); + CUDA_CHECK_RETURN(cudaPeekAtLastError()); + break; + case MOMENTUM: + case RMSPROP: + case ADAGRAD: + blocks = n/BLOCKSIZE_1STATE; + blocks = n % BLOCKSIZE_1STATE == 0 ? blocks : blocks + 1; + kOptimizerStatic8bit1StateBlockwise<T, OPTIMIZER, BLOCKSIZE_1STATE, NUM_1STATE><<<blocks, BLOCKSIZE_1STATE/NUM_1STATE>>>(p, g, state1, beta1, beta2, eps, step, lr, + quantiles1, absmax1, weight_decay, gnorm_scale, skip_zeros, n); + CUDA_CHECK_RETURN(cudaPeekAtLastError()); + break; + } } -template<typename T> -void percentileClipping(T *g, float *gnorm_vec, int step, const int n) { - int blocks = n / 2048; - blocks = n % 2048 == 0 ? blocks : blocks + 1; - CUDA_CHECK_RETURN(cudaMemset(&gnorm_vec[step % 100], 0, 1 * sizeof(float))); - kPercentileClipping < T, 2048, 4 ><<<blocks, 512>>>(g, gnorm_vec, step, n); - CUDA_CHECK_RETURN(cudaPeekAtLastError()); + +template<typename T> void percentileClipping(T * g, float *gnorm_vec, int step, const int n) +{ + int blocks = n/2048; + blocks = n % 2048 == 0 ? blocks : blocks + 1; + CUDA_CHECK_RETURN(cudaMemset(&gnorm_vec[step % 100], 0, 1*sizeof(float))); + kPercentileClipping<T, 2048, 4><<<blocks, 512>>>(g, gnorm_vec, step, n); + CUDA_CHECK_RETURN(cudaPeekAtLastError()); } @@ -195,23 +191,13 @@ void percentileClipping(T *g, float *gnorm_vec, int step, const int n) { //============================================================== template void estimateQuantiles(half *A, float *code, float offset, int n); - template void estimateQuantiles(float *A, float *code, float offset, int n); -template void -quantizeBlockwise<half, 0>(float *code, half *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n); - -template void -quantizeBlockwise<float, 0>(float *code, float *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n); - -template void -quantizeBlockwise<half, 1>(float *code, half *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n); - -template void -quantizeBlockwise<float, 1>(float *code, float *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n); - +template void quantizeBlockwise<half, 0>(float * code, half *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n); +template void quantizeBlockwise<float, 0>(float * code, float *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n); +template void quantizeBlockwise<half, 1>(float * code, half *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n); +template void quantizeBlockwise<float, 1>(float * code, float *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n); template void dequantizeBlockwise<half>(float *code, unsigned char *A, float *absmax, half *out, int blocksize, const int n); - template void dequantizeBlockwise<float>(float *code, unsigned char *A, float *absmax, float *out, int blocksize, const int n); #define MAKE_optimizer32bit(name, gtype) \ @@ -221,19 +207,12 @@ template void optimizer32bit<gtype, name>(gtype* g, gtype* p, \ const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n); MAKE_optimizer32bit(ADAM, half) - MAKE_optimizer32bit(ADAM, float) - MAKE_optimizer32bit(MOMENTUM, half) - MAKE_optimizer32bit(MOMENTUM, float) - MAKE_optimizer32bit(RMSPROP, half) - MAKE_optimizer32bit(RMSPROP, float) - MAKE_optimizer32bit(ADAGRAD, half) - MAKE_optimizer32bit(ADAGRAD, float) #define MAKE_optimizerStatic8bit(name, gtype) \ @@ -246,17 +225,11 @@ template void optimizerStatic8bit<gtype, name>(gtype* p, gtype* g, unsigned char float weight_decay, \ const float gnorm_scale, int n); \ - MAKE_optimizerStatic8bit(ADAM, half) - MAKE_optimizerStatic8bit(ADAM, float) - MAKE_optimizerStatic8bit(MOMENTUM, half) - MAKE_optimizerStatic8bit(MOMENTUM, float) - MAKE_optimizerStatic8bit(RMSPROP, half) - MAKE_optimizerStatic8bit(RMSPROP, float) #define MAKE_optimizerStatic8bitBlockwise(gtype, optim_name) \ @@ -264,23 +237,14 @@ template void optimizerStatic8bitBlockwise<gtype, optim_name>(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); \ - MAKE_optimizerStatic8bitBlockwise(half, ADAM); - MAKE_optimizerStatic8bitBlockwise(float, ADAM); - MAKE_optimizerStatic8bitBlockwise(half, MOMENTUM); - MAKE_optimizerStatic8bitBlockwise(float, MOMENTUM); - MAKE_optimizerStatic8bitBlockwise(half, RMSPROP); - MAKE_optimizerStatic8bitBlockwise(float, RMSPROP); - MAKE_optimizerStatic8bitBlockwise(half, ADAGRAD); - MAKE_optimizerStatic8bitBlockwise(float, ADAGRAD); -template void percentileClipping(float *g, float *gnorm_vec, int step, const int n); - -template void percentileClipping(half *g, float *gnorm_vec, int step, const int n); +template void percentileClipping(float * g, float *gnorm_vec, int step, const int n); +template void percentileClipping(half * g, float *gnorm_vec, int step, const int n); |