Added additional blocksizes: {64, 128, 256}.

2 files changed, 26 insertions, 2 deletions

diff --git a/csrc/kernels.cu b/csrc/kernels.cu
index 4c750d1..29f266a 100644
--- a/csrc/kernels.cu
+++ b/csrc/kernels.cu
@@ -454,8 +454,8 @@ __global__ void kQuantizeBlockwise(float * code, T * __restrict__ const A, float
   __shared__ float smem_code[256];
   __shared__ float smem_absmax_value[1];
 
-  if(threadIdx.x < 256)
-    smem_code[threadIdx.x] = code[threadIdx.x];
+  for(int i = threadIdx.x; i < 256; i+=blockDim.x)
+    smem_code[i] = code[i];
 
   for (unsigned int i = base_idx; i < n_full; i += gridDim.x*BLOCK_SIZE)
   {
@@ -2799,6 +2799,12 @@ template __global__ void kQuantizeBlockwise<half, 1024, 4, 0>(float * code, half
 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 kQuantizeBlockwise<half, 256, 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, 256, 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 kQuantizeBlockwise<half, 128, 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, 128, 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 kQuantizeBlockwise<half, 64, 1, 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, 64, 1, 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 * A, float * absmax, half *out, const int n);
 template __global__ void kDequantizeBlockwise<float, 4096, 1024, 4>(float *code, unsigned char * A, float * absmax, float *out, const int n);
@@ -2808,6 +2814,12 @@ template __global__ void kDequantizeBlockwise<half, 1024, 256, 4>(float *code, u
 template __global__ void kDequantizeBlockwise<float, 1024, 256, 4>(float *code, unsigned char * A, float * absmax, float *out, const int n);
 template __global__ void kDequantizeBlockwise<half, 512, 256, 2>(float *code, unsigned char * A, float * absmax, half *out, const int n);
 template __global__ void kDequantizeBlockwise<float, 512, 256, 2>(float *code, unsigned char * A, float * absmax, float *out, const int n);
+template __global__ void kDequantizeBlockwise<half, 256, 128, 2>(float *code, unsigned char * A, float * absmax, half *out, const int n);
+template __global__ void kDequantizeBlockwise<float, 256, 128, 2>(float *code, unsigned char * A, float * absmax, float *out, const int n);
+template __global__ void kDequantizeBlockwise<half, 128, 64, 2>(float *code, unsigned char * A, float * absmax, half *out, const int n);
+template __global__ void kDequantizeBlockwise<float, 128, 64, 2>(float *code, unsigned char * A, float * absmax, float *out, const int n);
+template __global__ void kDequantizeBlockwise<half, 64, 64, 1>(float *code, unsigned char * A, float * absmax, half *out, const int n);
+template __global__ void kDequantizeBlockwise<float, 64, 64, 1>(float *code, unsigned char * A, float * absmax, float *out, const int n);
 
 
 
diff --git a/csrc/ops.cu b/csrc/ops.cu
index b121fc2..30079e6 100644
--- a/csrc/ops.cu
+++ b/csrc/ops.cu
@@ -65,6 +65,12 @@ template <typename T, int STOCHASTIC> void quantizeBlockwise(float * code, T *A,
     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);
+  else if(blocksize == 256)
+    kQuantizeBlockwise<T, 256, 2, 0><<<num_blocks, 128>>>(code, A, absmax, out, rand, rand_offset, n);
+  else if(blocksize == 128)
+    kQuantizeBlockwise<T, 128, 2, 0><<<num_blocks, 64>>>(code, A, absmax, out, rand, rand_offset, n);
+  else if(blocksize == 64)
+    kQuantizeBlockwise<T, 64, 1, 0><<<num_blocks, 64>>>(code, A, absmax, out, rand, rand_offset, n);
 
 
   CUDA_CHECK_RETURN(cudaPeekAtLastError());
@@ -82,6 +88,12 @@ template<typename T> void dequantizeBlockwise(float *code, unsigned char *A, flo
     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);
+  else if(blocksize == 256)
+    kDequantizeBlockwise<T, 256, 128, 2><<<num_blocks, 256/2>>>(code, A, absmax, out, n);
+  else if(blocksize == 128)
+    kDequantizeBlockwise<T, 128, 64, 2><<<num_blocks, 128/2>>>(code, A, absmax, out, n);
+  else if(blocksize == 64)
+    kDequantizeBlockwise<T, 64, 64, 1><<<num_blocks, 64/1>>>(code, A, absmax, out, n);
 
   CUDA_CHECK_RETURN(cudaPeekAtLastError());
 }