Initial plumbing for skip_zeros.

5 files changed, 33 insertions, 32 deletions

diff --git a/csrc/kernels.cu b/csrc/kernels.cu
index 66a2c99..f8f7b62 100644
--- a/csrc/kernels.cu
+++ b/csrc/kernels.cu
@@ -654,7 +654,7 @@ __launch_bounds__(TH, 1)
 __global__ void kOptimizer32bit2State(T* g, T* p, 
                 float* state1, float* state2, float *unorm, const float max_unorm, const 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, const int n)
+                const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n)
 {
 
   const int n_full = ((TH*NUM_PER_THREAD)*(n/(TH*NUM_PER_THREAD))) + (n % (TH*NUM_PER_THREAD) == 0 ? 0 : (TH*NUM_PER_THREAD));
@@ -809,7 +809,7 @@ __launch_bounds__(TH, 1)
 __global__ void kOptimizer32bit1State(T *g, T *p, 
                 float *state1, float *unorm, const float max_unorm, const float param_norm,
                 const float beta1, const float eps, const float weight_decay,
-                const int step, const float lr, const float gnorm_scale, const int n)
+                const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n)
 {
 
   const int n_full = ((TH*NUM_PER_THREAD)*(n/(TH*NUM_PER_THREAD))) + (n % (TH*NUM_PER_THREAD) == 0 ? 0 : (TH*NUM_PER_THREAD));
@@ -1383,7 +1383,7 @@ kOptimizerStatic8bit2StateBlockwise(T* p, T* __restrict__ const g, unsigned char
                 float* __restrict__ const quantiles1, float* __restrict__ const quantiles2,
                 float* absmax1, float* absmax2, 
                 float weight_decay,
-                const float gnorm_scale, const int n)
+                const float gnorm_scale, const bool skip_zeros, const int n)
 {
 
     //const int n_full = n + (n%BLOCK_SIZE);
@@ -1555,7 +1555,7 @@ kOptimizerStatic8bit1StateBlockwise(T* p, T* __restrict__ const g, unsigned char
                 float* __restrict__ const quantiles1,
                 float* absmax1,
                 float weight_decay,
-                const float gnorm_scale, const int n)
+                const float gnorm_scale, const bool skip_zeros, const int n)
 {
 
     //const int n_full = n + (n%BLOCK_SIZE);
@@ -1723,7 +1723,7 @@ MAKE_PreconditionOptimizer32bit1State(RMSPROP, float)
 
 #define MAKE_Optimizer32bit1State(oname, gtype) \
 template __global__ void kOptimizer32bit1State<gtype, oname>(gtype* g, gtype* p, float* state1, float *unorm, const float max_unorm, const float param_norm, \
-    const float beta1, const float eps, const float weight_decay,const int step, const float lr, const float gnorm_scale, const int n); \
+    const float beta1, const float eps, const float weight_decay,const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n); \
 
 MAKE_Optimizer32bit1State(MOMENTUM, half)
 MAKE_Optimizer32bit1State(MOMENTUM, float)
@@ -1740,9 +1740,9 @@ MAKE_PreconditionOptimizer32bit2State(ADAM, half)
 MAKE_PreconditionOptimizer32bit2State(ADAM, float)
 
 template __global__ void kOptimizer32bit2State<half, ADAM>(half* g, half* p, float* state1, float* state2, float *unorm, const float max_unorm, const 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, const int n);
+    const float beta1, const float beta2, const float eps, const float weight_decay,const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n);
 template __global__ void kOptimizer32bit2State<float, ADAM>(float* g, float* p, float* state1, float* state2, float *unorm, const float max_unorm, const 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, const int n);
+    const float beta1, const float beta2, const float eps, const float weight_decay,const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n);
 
 #define MAKE_PreconditionStatic8bit1State(oname, gtype) \
 template __global__ void kPreconditionOptimizerStatic8bit1State<gtype, oname>(gtype* p, gtype* __restrict__ const g, unsigned char*__restrict__  const state1,  \
@@ -1825,7 +1825,7 @@ template __global__ void kOptimizerStatic8bit2StateBlockwise<gtype, oname, block
                 float* __restrict__ const quantiles1, float* __restrict__ const quantiles2, \
                 float* absmax1, float* absmax2,  \
                 float weight_decay, \
-                const float gnorm_scale, const int n); \
+                const float gnorm_scale, const bool skip_zeros, const int n); \
 
 MAKE_OptimizerStatic8bit2StateBlockwise(ADAM, float, 2048, 8)
 MAKE_OptimizerStatic8bit2StateBlockwise(ADAM, half, 2048, 8)
@@ -1838,7 +1838,7 @@ template __global__ void kOptimizerStatic8bit1StateBlockwise<gtype, oname, block
                 float* __restrict__ const quantiles1, \
                 float* absmax1, \
                 float weight_decay, \
-                const float gnorm_scale, const int n); \
+                const float gnorm_scale, const bool skip_zeros, const int n); \
 
 MAKE_OptimizerStatic8bit1StateBlockwise(MOMENTUM, float, 2048, 8)
 MAKE_OptimizerStatic8bit1StateBlockwise(MOMENTUM, half, 2048, 8)
diff --git a/csrc/kernels.cuh b/csrc/kernels.cuh
index 06ae1e4..0a3676c 100644
--- a/csrc/kernels.cuh
+++ b/csrc/kernels.cuh
@@ -27,7 +27,7 @@ template<typename T, int OPTIMIZER>
 __global__ void kOptimizer32bit2State(T* g, T* p, 
                 float* state1, float* state2, float *unorm, const float max_unorm, const 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, const int n);
+                const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n);
 
 template<typename T, int OPTIMIZER, int BLOCK_SIZE, int NUM_VALS>
 __global__ void kPreconditionOptimizer32bit1State(T* g, T* p, 
@@ -39,7 +39,7 @@ template<typename T, int OPTIMIZER>
 __global__ void kOptimizer32bit1State(T* g, T* p, 
                 float* state1,  float *unorm, const float max_unorm, const float param_norm,
                 const float beta1, const float eps, const float weight_decay,
-                const int step, const float lr, const float gnorm_scale, const int n);
+                const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n);
 
 template<typename T, int OPTIMIZER>
 __global__ void
@@ -90,7 +90,7 @@ template<typename T, int OPTIMIZER, int BLOCK_SIZE, int N_PER_TH> __global__ voi
 		T* p, T* __restrict__ const g, unsigned char* state1, unsigned char* state2,
                 const float beta1, const float beta2, const float eps, const int step, const float lr,
                 float* __restrict__ const quantiles1, float* __restrict__ const quantiles2,
-                float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, const int n);
+                float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, const bool skip_zeros, const int n);
 
 template<typename T, int OPTIMIZER, int BLOCK_SIZE, int N_PER_TH> __global__ void kOptimizerStatic8bit1StateBlockwise(
 		T* p, T* __restrict__ const g, unsigned char* state1,
@@ -99,7 +99,7 @@ template<typename T, int OPTIMIZER, int BLOCK_SIZE, int N_PER_TH> __global__ voi
                 float* __restrict__ const quantiles1,
                 float* absmax1,
                 float weight_decay,
-                const float gnorm_scale, const int n);
+                const float gnorm_scale, const bool skip_zeros, const int n);
 
 
 template<typename T, int BLOCK_SIZE, int NUM_VALS> __global__ void kPercentileClipping(T * __restrict__ g, float *gnorm_vec, int step, const int n);
diff --git a/csrc/ops.cu b/csrc/ops.cu
index d460ab1..182d6e6 100644
--- a/csrc/ops.cu
+++ b/csrc/ops.cu
@@ -181,7 +181,7 @@ template<typename T> void dequantizeBlockwise(float *code, unsigned char *A, flo
 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, const int n)
+                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;
@@ -194,7 +194,7 @@ template<typename T, int OPTIMIZER> void optimizer32bit(T* g, T* p,
         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, n);
+			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:
@@ -206,7 +206,7 @@ template<typename T, int OPTIMIZER> void optimizer32bit(T* g, T* p,
         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, n);
+			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;
 	}
@@ -259,7 +259,7 @@ template<typename T, int OPTIMIZER> void optimizerStatic8bit(T* p, T* g,
 
 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, int n)
+                float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, bool skip_zeros, int n)
 {
 
 	int blocks = 0;
@@ -269,7 +269,7 @@ template<typename T, int OPTIMIZER> void optimizerStatic8bitBlockwise(T* p, T* g
 			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, n);
+																														quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, skip_zeros, n);
 			CUDA_CHECK_RETURN(cudaPeekAtLastError());
 		break;
 		case MOMENTUM:
@@ -277,7 +277,7 @@ template<typename T, int OPTIMIZER> void optimizerStatic8bitBlockwise(T* p, T* g
 			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, n);
+																														quantiles1, absmax1, weight_decay, gnorm_scale, skip_zeros, n);
 			CUDA_CHECK_RETURN(cudaPeekAtLastError());
 		break;
 	}
@@ -313,7 +313,7 @@ template void dequantizeBlockwise<float>(float *code, unsigned char *A, float *a
 template void optimizer32bit<gtype, name>(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, const int n);
+                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)
@@ -342,7 +342,7 @@ MAKE_optimizerStatic8bit(RMSPROP, float)
 #define MAKE_optimizerStatic8bitBlockwise(gtype, optim_name) \
 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, int n); \
+                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);
diff --git a/csrc/ops.cuh b/csrc/ops.cuh
index e6033cb..465b4a4 100644
--- a/csrc/ops.cuh
+++ b/csrc/ops.cuh
@@ -49,7 +49,7 @@ template<typename T> void dequantizeBlockwise(float *code, unsigned char *A, flo
 template<typename T, int OPTIMIZER> void optimizer32bit(T* g, T* p, 
                 float* state1, float* state2, float *unorm, float max_unorm, float param_norm,
                 float beta1, float beta2, float eps, float weight_decay,
-                int step, float lr, const float gnorm_scale, int n);
+                int step, float lr, const float gnorm_scale, bool skip_zeros, int n);
 
 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,
@@ -62,7 +62,8 @@ template<typename T, int OPTIMIZER> void optimizerStatic8bit(T* p, T* g, unsigne
 
 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, int n);
+                float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, 
+								bool skip_zeros, int n);
 
 template<typename T> void percentileClipping(T * g, float *gnorm_vec, int step, const int n);
 
diff --git a/csrc/pythonInterface.c b/csrc/pythonInterface.c
index eacb849..67bf2e5 100644
--- a/csrc/pythonInterface.c
+++ b/csrc/pythonInterface.c
@@ -20,8 +20,8 @@ void estimateQuantiles_fp16(half *A, float *code, float offset, int n){ estimate
 void fname##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, float gnorm_scale, const int n) \
-{ optimizer32bit<gtype, oname>(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, n); } \
+               const int step, const float lr, float gnorm_scale, bool skip_zeros, const int n) \
+{ optimizer32bit<gtype, oname>(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n); } \
 
 MAKE_FUNC32(momentum, MOMENTUM, float, 32)
 MAKE_FUNC32(momentum, MOMENTUM, half, 16)
@@ -53,8 +53,8 @@ MAKE_FUNC8(rmsprop, RMSPROP, half, 16)
 #define MAKE_BLOCKWISE8(fname, optim_name, gtype, gbits) \
 void 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, int n)\
-{	optimizerStatic8bitBlockwise<gtype, optim_name>(p, g, state1, state2, beta1, beta2, eps, step, lr, quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, n); }\
+                float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, bool skip_zeros, int n)\
+{	optimizerStatic8bitBlockwise<gtype, optim_name>(p, g, state1, state2, beta1, beta2, eps, step, lr, quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, skip_zeros, n); }\
 
 MAKE_BLOCKWISE8(adam, ADAM, half, 16)
 MAKE_BLOCKWISE8(adam, ADAM, float, 32)
@@ -93,8 +93,8 @@ extern "C"
 	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, 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, n); } \
+								 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)
@@ -110,7 +110,7 @@ extern "C"
                 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) \
+                float weight_decay, float gnorm_scale, bool skip_zeros, 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); \
@@ -126,8 +126,8 @@ extern "C"
   #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, int n) \
-  {	fname##_8bit_blockwise_fp##gbits(p, g, state1, state2, beta1, beta2, eps, step, lr, quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, n); } \
+                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)