Added blocksizes 2048, 1024, and 512 to blockwise quant.

4 files changed, 48 insertions, 21 deletions

diff --git a/csrc/kernels.cu b/csrc/kernels.cu
index f01b4e1..9d9653c 100644
--- a/csrc/kernels.cu
+++ b/csrc/kernels.cu
@@ -428,16 +428,16 @@ __global__ void kQuantize(float * code, float * __restrict__ const A, unsigned c
 }
 
 template<typename T, int BLOCK_SIZE, int NUM_PER_TH, int STOCHASTIC>
-__launch_bounds__(TH, 4)
+//__launch_bounds__(TH, 4)
 __global__ void kQuantizeBlockwise(float * code, T * __restrict__ const A, float *absmax, unsigned char *out, float * __restrict__ const rand, const int rand_offset, const int n)
 {
   const int n_full = gridDim.x * BLOCK_SIZE;
   int valid_items = 0;
   const int base_idx = (blockIdx.x * BLOCK_SIZE);
 
-  T vals[NUM];
-  float rand_vals[NUM];
-  unsigned char qvals[NUM];
+  T vals[NUM_PER_TH];
+  float rand_vals[NUM_PER_TH];
+  unsigned char qvals[NUM_PER_TH];
   //float local_abs_max = -FLT_MAX;
   float local_abs_max = 0.0f;
   int local_rand_idx = 0;
@@ -517,8 +517,8 @@ __global__ void kDequantizeBlockwise(float *code, unsigned char * __restrict__ c
   int valid_items = 0;
   const int base_idx = (blockIdx.x * BLOCK_SIZE);
 
-  T vals[NUM];
-  unsigned char qvals[NUM];
+  T vals[NUM_PER_TH];
+  unsigned char qvals[NUM_PER_TH];
   float local_abs_max = -FLT_MAX;
 
   typedef cub::BlockLoad<unsigned char, THREADS, NUM_PER_TH, cub::BLOCK_LOAD_WARP_TRANSPOSE> LoadChar;
@@ -2791,11 +2791,21 @@ template __global__ void kQuantizeBlockwise<half, 4096, 4, 0>(float * code, half
 template __global__ void kQuantizeBlockwise<float, 4096, 4, 0>(float * code, float * __restrict__ const A, float *absmax, unsigned char *out, float * __restrict__ const rand, const int rand_offset, const int n);
 template __global__ void kQuantizeBlockwise<half, 4096, 4, 1>(float * code, half * __restrict__ const A, float *absmax, unsigned char *out, float * __restrict__ const rand, const int rand_offset, const int n);
 template __global__ void kQuantizeBlockwise<float, 4096, 4, 1>(float * code, float * __restrict__ const A, float *absmax, unsigned char *out, float * __restrict__ const rand, const int rand_offset, const int n);
+template __global__ void kQuantizeBlockwise<half, 2048, 4, 0>(float * code, half * __restrict__ const A, float *absmax, unsigned char *out, float * __restrict__ const rand, const int rand_offset, const int n);
+template __global__ void kQuantizeBlockwise<float, 2048, 4, 0>(float * code, float * __restrict__ const A, float *absmax, unsigned char *out, float * __restrict__ const rand, const int rand_offset, const int n);
+template __global__ void kQuantizeBlockwise<half, 1024, 4, 0>(float * code, half * __restrict__ const A, float *absmax, unsigned char *out, float * __restrict__ const rand, const int rand_offset, const int n);
+template __global__ void kQuantizeBlockwise<float, 1024, 4, 0>(float * code, float * __restrict__ const A, float *absmax, unsigned char *out, float * __restrict__ const rand, const int rand_offset, const int n);
+template __global__ void kQuantizeBlockwise<half, 512, 2, 0>(float * code, half * __restrict__ const A, float *absmax, unsigned char *out, float * __restrict__ const rand, const int rand_offset, const int n);
+template __global__ void kQuantizeBlockwise<float, 512, 2, 0>(float * code, float * __restrict__ const A, float *absmax, unsigned char *out, float * __restrict__ const rand, const int rand_offset, const int n);
 
 template __global__ void kDequantizeBlockwise<half, 4096, 1024, 4>(float *code, unsigned char * __restrict__ const A, float * __restrict__ const absmax, half *out, const int n);
 template __global__ void kDequantizeBlockwise<float, 4096, 1024, 4>(float *code, unsigned char * __restrict__ const A, float * __restrict__ const absmax, float *out, const int n);
 template __global__ void kDequantizeBlockwise<half, 2048, 512, 4>(float *code, unsigned char * __restrict__ const A, float * __restrict__ const absmax, half *out, const int n);
 template __global__ void kDequantizeBlockwise<float, 2048, 512, 4>(float *code, unsigned char * __restrict__ const A, float * __restrict__ const absmax, float *out, const int n);
+template __global__ void kDequantizeBlockwise<half, 1024, 256, 4>(float *code, unsigned char * __restrict__ const A, float * __restrict__ const absmax, half *out, const int n);
+template __global__ void kDequantizeBlockwise<float, 1024, 256, 4>(float *code, unsigned char * __restrict__ const A, float * __restrict__ const absmax, float *out, const int n);
+template __global__ void kDequantizeBlockwise<half, 512, 256, 2>(float *code, unsigned char * __restrict__ const A, float * __restrict__ const absmax, half *out, const int n);
+template __global__ void kDequantizeBlockwise<float, 512, 256, 2>(float *code, unsigned char * __restrict__ const A, float * __restrict__ const absmax, float *out, const int n);
 
 
 
diff --git a/csrc/ops.cu b/csrc/ops.cu
index e49c94b..b121fc2 100644
--- a/csrc/ops.cu
+++ b/csrc/ops.cu
@@ -50,11 +50,23 @@ 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)
+template <typename T, int STOCHASTIC> void quantizeBlockwise(float * code, T *A, float *absmax, unsigned char *out, float *rand, int rand_offset, int blocksize, const int n)
 {
-  int num_blocks = n/4096;
-  num_blocks = n % 4096 == 0 ? num_blocks : num_blocks + 1;
-  kQuantizeBlockwise<T, 4096, 4, STOCHASTIC><<<num_blocks, 1024>>>(code, A, absmax, out, rand, rand_offset, n);
+  int num_blocks = n/blocksize;
+  num_blocks = n % blocksize == 0 ? num_blocks : num_blocks + 1;
+  if(STOCHASTIC == 1)
+    assert(blocksize == 4096);
+
+  if(blocksize == 4096)
+    kQuantizeBlockwise<T, 4096, 4, STOCHASTIC><<<num_blocks, 1024>>>(code, A, absmax, out, rand, rand_offset, n);
+  else if(blocksize == 2048)
+    kQuantizeBlockwise<T, 2048, 4, 0><<<num_blocks, 512>>>(code, A, absmax, out, rand, rand_offset, n);
+  else if(blocksize == 1024)
+    kQuantizeBlockwise<T, 1024, 4, 0><<<num_blocks, 256>>>(code, A, absmax, out, rand, rand_offset, n);
+  else if(blocksize == 512)
+    kQuantizeBlockwise<T, 512, 2, 0><<<num_blocks, 256>>>(code, A, absmax, out, rand, rand_offset, n);
+
+
   CUDA_CHECK_RETURN(cudaPeekAtLastError());
 }
 
@@ -66,6 +78,11 @@ template<typename T> void dequantizeBlockwise(float *code, unsigned char *A, flo
     kDequantizeBlockwise<T, 4096, 1024, 4><<<num_blocks, 4096/4>>>(code, A, absmax, out, n);
   else if(blocksize == 2048)
     kDequantizeBlockwise<T, 2048, 512, 4><<<num_blocks, 2048/4>>>(code, A, absmax, out, n);
+  else if(blocksize == 1024)
+    kDequantizeBlockwise<T, 1024, 256, 4><<<num_blocks, 1024/4>>>(code, A, absmax, out, n);
+  else if(blocksize == 512)
+    kDequantizeBlockwise<T, 512, 256, 2><<<num_blocks, 512/2>>>(code, A, absmax, out, n);
+
   CUDA_CHECK_RETURN(cudaPeekAtLastError());
 }
 
@@ -659,10 +676,10 @@ template void transformRowToFormat<COL_AMPERE, 1>(char * A, char *out, int rows,
 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, int blocksize, const int n);
+template void quantizeBlockwise<float, 0>(float * code, float *A, float *absmax, unsigned char *out, float* rand, int rand_offset, int blocksize, const int n);
+template void quantizeBlockwise<half, 1>(float * code, half *A, float *absmax, unsigned char *out, float* rand, int rand_offset, int blocksize, const int n);
+template void quantizeBlockwise<float, 1>(float * code, float *A, float *absmax, unsigned char *out, float* rand, int rand_offset, int blocksize, 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);
 
diff --git a/csrc/ops.cuh b/csrc/ops.cuh
index acfdb06..66e3843 100644
--- a/csrc/ops.cuh
+++ b/csrc/ops.cuh
@@ -128,7 +128,7 @@ template <typename T> void estimateQuantiles(T *A, float *code, float offset, in
 
 void quantize(float *code, float *A, unsigned char *out, int n);
 void dequantize(float *code, unsigned char *A, float *out, int n);
-template <typename T, int STOCHASTIC> void quantizeBlockwise(float * code, T *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n);
+template <typename T, int STOCHASTIC> void quantizeBlockwise(float * code, T *A, float *absmax, unsigned char *out, float* rand, int rand_offset, int blocksize, const int n);
 template<typename T> void dequantizeBlockwise(float *code, unsigned char *A, float *absmax, T *out, int block_size, const int n);
 
 template<typename T, int OPTIMIZER> void optimizer32bit(T* g, T* p, 
diff --git a/csrc/pythonInterface.c b/csrc/pythonInterface.c
index 58e26a9..5bac30e 100644
--- a/csrc/pythonInterface.c
+++ b/csrc/pythonInterface.c
@@ -75,10 +75,10 @@ MAKE_BLOCKWISE8(adagrad, ADAGRAD, float, 32)
 void percentileClipping_g32(float * g, float *gnorm_vec, int step, const int n){ percentileClipping<float>(g, gnorm_vec, step, n); }
 void percentileClipping_g16(half * g, float *gnorm_vec, int step, const int n){ percentileClipping<half>(g, gnorm_vec, step, n); }
 
-void quantizeBlockwise_fp16(float * code, half *A, float *absmax, unsigned char *out, const int n){ quantizeBlockwise<half, 0>(code, A, absmax, out, NULL, 0, n); }
-void quantizeBlockwise_fp32(float * code, float *A, float *absmax, unsigned char *out, const int n){ quantizeBlockwise<float, 0>(code, A, absmax, out, NULL, 0, n); }
-void quantizeBlockwise_stochastic_fp16(float * code, half *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n){ quantizeBlockwise<half, 1>(code, A, absmax, out, rand, rand_offset, n); }
-void quantizeBlockwise_stochastic_fp32(float * code, float *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n){ quantizeBlockwise<float, 1>(code, A, absmax, out, rand, rand_offset, n); }
+void quantizeBlockwise_fp16(float * code, half *A, float *absmax, unsigned char *out, int blocksize, const int n){ quantizeBlockwise<half, 0>(code, A, absmax, out, NULL, 0, blocksize, n); }
+void quantizeBlockwise_fp32(float * code, float *A, float *absmax, unsigned char *out, int blocksize, const int n){ quantizeBlockwise<float, 0>(code, A, absmax, out, NULL, 0, blocksize, n); }
+void quantizeBlockwise_stochastic_fp16(float * code, half *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n){ quantizeBlockwise<half, 1>(code, A, absmax, out, rand, rand_offset, 4096, n); }
+void quantizeBlockwise_stochastic_fp32(float * code, float *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n){ quantizeBlockwise<float, 1>(code, A, absmax, out, rand, rand_offset, 4096, n); }
 
 void dequantizeBlockwise_fp16(float *code, unsigned char *A, float *absmax, half *out, int blocksize, const int n){ dequantizeBlockwise<half>(code, A, absmax, out, blocksize, n); } \
 void dequantizeBlockwise_fp32(float *code, unsigned char *A, float *absmax, float *out, int blocksize, const int n){ dequantizeBlockwise<float>(code, A, absmax, out, blocksize, n); }
@@ -140,8 +140,8 @@ extern "C"
 	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_fp16(float * code, half *A, float *absmax, unsigned char *out, int blocksize, const int n){ quantizeBlockwise_fp16(code, A, absmax, out, blocksize, n); }
+  void cquantize_blockwise_fp32(float * code, float *A, float *absmax, unsigned char *out, int blocksize, const int n){ quantizeBlockwise_fp32(code, A, absmax, out, blocksize, 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); }