Merge branch 'extract_outliers' into debug

5 files changed, 108 insertions, 6 deletions

diff --git a/csrc/kernels.cu b/csrc/kernels.cu
index 6eca3aa..d4eb56c 100644
--- a/csrc/kernels.cu
+++ b/csrc/kernels.cu
@@ -2591,16 +2591,82 @@ __global__ void kspmm_coo_very_sparse_naive(int *max_count, int *max_idx, int *o
   } 
 }
 
+template <int FORMAT> __global__ void kExtractOutliers(char *A, int *idx, char *out, int idx_size, int rowsA, int colsA, int tiledRowsA, int tiledColsA)
+{  
+	int local_colidx = idx[blockIdx.x];
+
+	if(FORMAT==COL_TURING)
+	{
+		// TURING FORMAT:
+		// 8*32 tiles with 4*4 subtiles
+		// the 8*32 subtile has first all 4*4 subtiles of even rows (max 4*4*8 = 128 elements)
+		// the subsequent 4*4 subtiles are for all odd rows if some rows columns are empty the values are zero
+		// the tile repeats again after the 8*32 tile in a major column order, meaning: (next 8 rows are A[8:16, 0:32])
+		// the next tile is the next 8 rows for the same 32 columns. Once all rows are finished, the column
+		// index increases by 32
+		// columns are grouped in increments of 4, meaning that one has the following rows and columns
+		// rows: [0 0 0 0, 2 2 2 2, 4 4 4 4, 6 6 6 6, 0 0 0 0 ...]
+		// cols: [0 1 2 3, 0 1 2 4, 0 1 2 3, 0 1 2 3, 4 5 6 7 ...]
+
+		// each thread reads 1 element = 1 row
+		for(int row = threadIdx.x; row < rowsA; row+= blockDim.x)
+		{
+			int offset_per_col_tile = ((rowsA+7)/8)*32*8;
+			int tile_offset_rows = (row/8)*32*8;
+			int tile_offset_cols = (local_colidx/32)*offset_per_col_tile;
+			int offset = 0;
+			int subtile_col_idx = local_colidx%32;
+			int subtile_row_idx = row % 8;
+			if(row % 2 == 1)
+				offset += 128 + (subtile_col_idx/4)*16 + (subtile_col_idx%4) + ((subtile_row_idx-1)*2);
+			else
+				// even
+				offset += 0   + (subtile_col_idx/4)*16 + (subtile_col_idx%4) + (subtile_row_idx*2);
+
+			offset += tile_offset_rows + tile_offset_cols;
+
+			char val = A[offset];
+
+			int out_idx = (row*idx_size) + blockIdx.x;
+			out[out_idx] = val;
+		}
+	}
+	else if(FORMAT == COL_AMPERE)
+	{
+
+		for(int row = threadIdx.x; row < rowsA; row+= blockDim.x)
+		{
+			// we got 32x32 tiles and we use the magic equation from the cublasLt doc to get the element
+			// within each tile.
+			int offset_per_col_tile = ((rowsA+31)/32)*32*32;
+			int tile_offset_rows = (row/32)*32*32;
+			int tile_offset_cols = (local_colidx/32)*offset_per_col_tile;
+			int subtile_col_idx = local_colidx%32;
+			int subtile_row_idx = row % 32;
+			// this magic is taken from the cublasLt doc (search for COL32)
+			int offset = (((subtile_row_idx%8)/2*4+subtile_row_idx/8)*2+subtile_row_idx%2)*32+subtile_col_idx;
+			offset += tile_offset_cols + tile_offset_rows;
+
+			char val = A[offset];
+			int out_idx = (row*idx_size) + blockIdx.x;
+			out[out_idx] = val;
+		}
+	}
+} 
+
 //==============================================================
 //                   TEMPLATE DEFINITIONS
 //==============================================================
 
-template __global__ void kspmm_coo_very_sparse_naive<half, 8, 16>(int *max_count, int *max_idx, int *offset_rowidx, int *rowidx, int *colidx, half *values, half *B, half *out, float * __restrict__ const dequant_stats, int nnz, int rowsA, int rowsB, int colsB);
-template __global__ void kspmm_coo_very_sparse_naive<half, 16, 16>(int *max_count, int *max_idx, int *offset_rowidx, int *rowidx, int *colidx, half *values, half *B, half *out, float * __restrict__ const dequant_stats, int nnz, int rowsA, int rowsB, int colsB);
-template __global__ void kspmm_coo_very_sparse_naive<half, 32, 16>(int *max_count, int *max_idx, int *offset_rowidx, int *rowidx, int *colidx, half *values, half *B, half *out, float * __restrict__ const dequant_stats, int nnz, int rowsA, int rowsB, int colsB);
-template __global__ void kspmm_coo_very_sparse_naive<signed char, 8, 8>(int *max_count, int *max_idx, int *offset_rowidx, int *rowidx, int *colidx, half *values, signed char *B, half *out, float * __restrict__ const dequant_stats, int nnz, int rowsA, int rowsB, int colsB);
-template __global__ void kspmm_coo_very_sparse_naive<signed char, 16, 8>(int *max_count, int *max_idx, int *offset_rowidx, int *rowidx, int *colidx, half *values, signed char *B, half *out, float * __restrict__ const dequant_stats, int nnz, int rowsA, int rowsB, int colsB);
-template __global__ void kspmm_coo_very_sparse_naive<signed char, 32, 8>(int *max_count, int *max_idx, int *offset_rowidx, int *rowidx, int *colidx, half *values, signed char *B, half *out, float * __restrict__ const dequant_stats, int nnz, int rowsA, int rowsB, int colsB);
+template __global__ void kExtractOutliers<COL_TURING>(char *A, int *idx, char *out, int idx_size, int rowsA, int colsA, int tiledRowsA, int tiledColsA);
+template __global__ void kExtractOutliers<COL_AMPERE>(char *A, int *idx, char *out, int idx_size, int rowsA, int colsA, int tiledRowsA, int tiledColsA);
+
+template __global__ void kspmm_coo_very_sparse_naive<half, 8, 16>(int *max_count, int *max_idx, int *offset_rowidx, int *rowidx, int *colidx, half *values, half *B, half *out, float *dequant_stats, int nnz, int rowsA, int rowsB, int colsB);
+template __global__ void kspmm_coo_very_sparse_naive<half, 16, 16>(int *max_count, int *max_idx, int *offset_rowidx, int *rowidx, int *colidx, half *values, half *B, half *out, float *dequant_stats, int nnz, int rowsA, int rowsB, int colsB);
+template __global__ void kspmm_coo_very_sparse_naive<half, 32, 16>(int *max_count, int *max_idx, int *offset_rowidx, int *rowidx, int *colidx, half *values, half *B, half *out, float *dequant_stats, int nnz, int rowsA, int rowsB, int colsB);
+template __global__ void kspmm_coo_very_sparse_naive<signed char, 8, 8>(int *max_count, int *max_idx, int *offset_rowidx, int *rowidx, int *colidx, half *values, signed char *B, half *out, float *dequant_stats, int nnz, int rowsA, int rowsB, int colsB);
+template __global__ void kspmm_coo_very_sparse_naive<signed char, 16, 8>(int *max_count, int *max_idx, int *offset_rowidx, int *rowidx, int *colidx, half *values, signed char *B, half *out, float *dequant_stats, int nnz, int rowsA, int rowsB, int colsB);
+template __global__ void kspmm_coo_very_sparse_naive<signed char, 32, 8>(int *max_count, int *max_idx, int *offset_rowidx, int *rowidx, int *colidx, half *values, signed char *B, half *out, float *dequant_stats, int nnz, int rowsA, int rowsB, int colsB);
 
 template __global__ void kTransformRowToFormat<256, 8, 32, 32*8, 0, COL32>(char *__restrict__ const A, char *out, int rows, int cols, int tiledCols, int outRows, int outCols);
 template __global__ void kTransformRowToFormat<256, 8, 32, 32*8, 1, COL32>(char *__restrict__ const A, char *out, int rows, int cols, int tiledCols, int outRows, int outCols);
diff --git a/csrc/kernels.cuh b/csrc/kernels.cuh
index 4e65e96..2447494 100644
--- a/csrc/kernels.cuh
+++ b/csrc/kernels.cuh
@@ -118,6 +118,8 @@ template <int THREADS, int ITEMS_PER_THREAD, int TILE_ROWS, int TILE_COLS, int S
 
 template <int THREADS, int ITEMS_PER_THREAD, int TILE_ROWS, int TILE_COLS, int TRANSPOSE, int FORMAT> __global__ void kTransformRowToFormat(char *__restrict__ const A, char *out, int rows, int cols, int tiledCols, int outRows, int outCols);
 
+template <int FORMAT> __global__ void kExtractOutliers(char *A, int *idx, char *out, int idx_size, int rowsA, int colsA, int tiledRowsA, int tiledColsA);
+
 #endif
 
 
diff --git a/csrc/ops.cu b/csrc/ops.cu
index cfc9605..9b75e69 100644
--- a/csrc/ops.cu
+++ b/csrc/ops.cu
@@ -618,10 +618,36 @@ template <typename T, int BITS> void spmm_coo_very_sparse_naive(int *max_count,
   CUDA_CHECK_RETURN(cudaPeekAtLastError());
 }
 
+
+template <int FORMAT> void extractOutliers(char * A, int *idx, char *out, int idx_size, int rows, int cols)
+{
+  int threads = 256;
+  // we load 128 column values per warp
+  int tiledCols = tiledCols = fill_up_to_nearest_multiple(cols, 32);
+  int tiledRows = 0;
+
+	int num_blocks = idx_size;
+
+  if(FORMAT == COL_TURING)
+  {
+      tiledRows = fill_up_to_nearest_multiple(rows, 8);
+  }
+  else if(FORMAT == COL_AMPERE)
+  {
+      tiledRows = fill_up_to_nearest_multiple(rows, 32);
+	}
+
+  kExtractOutliers<FORMAT><<<num_blocks, threads>>>(A, idx, out, idx_size, rows, cols, tiledRows, tiledCols);
+  CUDA_CHECK_RETURN(cudaPeekAtLastError());
+}
+
 //==============================================================
 //                   TEMPLATE DEFINITIONS
 //==============================================================
 
+template void extractOutliers<COL_TURING>(char * A, int *idx, char *out, int idx_size, int rows, int cols);
+template void extractOutliers<COL_AMPERE>(char * A, int *idx, char *out, int idx_size, int rows, int cols);
+
 template void spmm_coo_very_sparse_naive<half, 16>(int *max_count, int *max_idx, int *offset_rowidx, int *rowidx, int *colidx, half *values, half *B, half *out, float *dequant_stats, int nnz_rows, int nnz, int rowsA, int rowsB, int colsB);
 template void spmm_coo_very_sparse_naive<signed char, 8>(int *max_count, int *max_idx, int *offset_rowidx, int *rowidx, int *colidx, half *values, signed char *B, half *out, float *dequant_stats, int nnz_rows, int nnz, int rowsA, int rowsB, int colsB);
 
diff --git a/csrc/ops.cuh b/csrc/ops.cuh
index 4e719df..4b09ecf 100644
--- a/csrc/ops.cuh
+++ b/csrc/ops.cuh
@@ -174,4 +174,6 @@ void spmm_coo(cusparseHandle_t handle, int *A_rowidx, int *A_colidx, half *A_val
 
 template <typename T, int BITS> void spmm_coo_very_sparse_naive(int *max_count, int *max_idx, int *offset_rowidx, int *rowidx, int *colidx, half *values, T *B, half *out, float *dequant_stats, int nnz_rows, int nnz, int rowsA, int rowsB, int colsB);
 
+template <int FORMAT> void extractOutliers(char * A, int *idx, char *out, int idx_size, int rows, int cols);
+
 #endif
diff --git a/csrc/pythonInterface.c b/csrc/pythonInterface.c
index a6a4b13..7356c11 100644
--- a/csrc/pythonInterface.c
+++ b/csrc/pythonInterface.c
@@ -105,6 +105,9 @@ void transform_row2turingT(char * A, char *out, int rows, int cols){ transformRo
 void transform_row2ampere(char * A, char *out, int rows, int cols){ transformRowToFormat<COL_AMPERE, 0>(A, out, rows, cols); }
 void transform_row2ampereT(char * A, char *out, int rows, int cols){ transformRowToFormat<COL_AMPERE, 1>(A, out, rows, cols); }
 
+void extractOutliers_turing(char * A, int *idx, char *out, int idx_size, int rows, int cols){ extractOutliers<COL_TURING>(A, idx, out, idx_size, rows, cols); }
+void extractOutliers_ampere(char * A, int *idx, char *out, int idx_size, int rows, int cols){ extractOutliers<COL_AMPERE>(A, idx, out, idx_size, rows, cols); }
+
  int igemmlt_turing_32(cublasLtHandle_t ltHandle, int m, int n, int k, const int8_t *A, const int8_t *B, void *C, float *row_scale, int lda, int ldb, int ldc)
 	{ return igemmlt<COL_TURING, 32, 0>(ltHandle, m, n, k, A, B, C, row_scale, lda, ldb, ldc); }
 
@@ -280,6 +283,9 @@ extern "C"
 	void cspmm_coo_very_sparse_naive_int8(int *max_count, int *max_idx, int *offset_rowidx, int *rowidx, int *colidx, half *values, signed char *B, half *out, float *dequant_stats, int nnz_rows, int nnz, int rowsA, int rowsB, int colsB)
 	{ spmm_coo_very_sparse_naive_int8(max_count, max_idx, offset_rowidx, rowidx, colidx, values, B, out, dequant_stats, nnz_rows, nnz, rowsA, rowsB, colsB); }
 
+	void cextractOutliers_turing(char * A, int *idx, char *out, int idx_size, int rows, int cols){ extractOutliers_turing(A, idx, out, idx_size, rows, cols); }
+	void cextractOutliers_ampere(char * A, int *idx, char *out, int idx_size, int rows, int cols){ extractOutliers_ampere(A, idx, out, idx_size, rows, cols); }
+
 #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); }