Skip to content

Commit 11e2e92

Browse files
matthewdouglasmatthewdouglas
committed
Cleanup obsolete C++/CUDA code
1 parent 0b04376 commit 11e2e92

File tree

4 files changed

+0
-496
lines changed

4 files changed

+0
-496
lines changed

‎csrc/kernels.cu‎

Lines changed: 0 additions & 334 deletions
Original file line numberDiff line numberDiff line change
@@ -2205,333 +2205,6 @@ __global__ void kdequant_mm_int32_fp16(
22052205
}
22062206
}
22072207

2208-
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)
2209-
{
2210-
2211-
// 0. Load data into 32*32 shared memory tiles
2212-
// 1. transpose / reorder in shared memory
2213-
// 2. store
2214-
2215-
// COL32 FORMAT:
2216-
// rows*32 tiles
2217-
2218-
// TURING FORMAT:
2219-
// 8*32 tiles with 4*4 subtiles
2220-
// the 8*32 subtile has first all 4*4 subtiles of even rows (max 4*4*4 = 64 elements)
2221-
// the subsequent 4*4 subtiles are for all odd rows if some rows columns are empty the values are zero
2222-
// the tile repeats again after the 8*32 tile in a major column order, meaning: (next 8 rows are A[8:16, 0:32])
2223-
// the next tile is the next 8 rows for the same 32 columns. Once all rows are finished, the column
2224-
// index increases by 32
2225-
2226-
// AMPERE FORMAT:
2227-
// 32*32 tiles with 8*32 subtiles. The rows are interleaved in pairs of two rows with offset of 8 between pairs of two rows:
2228-
// row idx (each number stands for 32 values): [0 1 8 9 16 17 24 25] [2 3 10 11 18 19 26 27]...
2229-
// the tiles are column-major ordered, so after 1024*1024 values we process: A[32:64, 0:32]
2230-
2231-
2232-
// To have efficient loads and stores if we transpose we need 128 consequitive bytes which at 1 byte are 128 values
2233-
// As such we need:
2234-
// at least 32*4 shared memory tiles for col32; preferably 32*32
2235-
// at least 32*6 shared memory tiles for col32_ampere: preferably 32*32
2236-
// at least 32*8 shared memory tiles for col4_turing: preferably 32*32
2237-
// for efficient loading of row major we need to load 128 elements and repeat this 32 items
2238-
// this would imply a 32x128 shared memory tile -> 4kb
2239-
// It is more efficient to have more than 1 warp, so with 64 threads we need 32x128 -> 8 kb
2240-
// we have 64k sharded mem per SM in Turing which is 8 blocks per SM which is 2*8 = 32 warps = 100% occupancy
2241-
// for turing and 50% for A100 and 75% for RTX 30s / A40 which is probably good enough
2242-
// register pressure should be low with: 8 registers from local memoryh per block and 64 registers per SM
2243-
//
2244-
// to make the shared memory work with that occupancy we might need to union the block loads/stores
2245-
2246-
// each block loads TILE_COLs columns and TILE_ROW rows
2247-
// after reading a tile the row counter increase by TILE_ROWS
2248-
// the col counter reset after reading TILE_COL elements
2249-
const int base_row = ((blockIdx.x*TILE_COLS)/tiledCols)*TILE_ROWS;
2250-
// col increases by TILE_SIZE for each block and wraps back to 0 after tiledCols is reached
2251-
const int base_col = (blockIdx.x*TILE_COLS) % tiledCols;
2252-
const int base_idx = (base_row*cols) + base_col;
2253-
2254-
// we load 128 bytes per warp with
2255-
// 32 rows for transposes that fill col32 types
2256-
// so that we can have contiguous stores
2257-
__shared__ char smem_data[32*33*ITEMS_PER_THREAD];
2258-
char local_data[ITEMS_PER_THREAD];
2259-
typedef cub::BlockExchange<char, THREADS, ITEMS_PER_THREAD> BlockExchange;
2260-
2261-
// we load row after row from the base_position
2262-
// Load data row by row
2263-
int warps = blockDim.x/32;
2264-
int warp_id = threadIdx.x/32;
2265-
int warp_lane = threadIdx.x % 32;
2266-
int offset = 0;
2267-
2268-
int smem_row = 0;
2269-
// each warp loads one row of 128 bytes
2270-
for(int row = warp_id; row < TILE_ROWS; row+=warps)
2271-
{
2272-
int i = base_idx + (row*cols);
2273-
// we load up to 128 bytes/items per load
2274-
int valid_items = cols - base_col > 32*ITEMS_PER_THREAD ? 32*ITEMS_PER_THREAD : cols - base_col;
2275-
2276-
// 0. Load data into 32*32 shared memory tiles
2277-
if(base_row + row < rows)
2278-
{
2279-
#pragma unroll ITEMS_PER_THREAD
2280-
for(int j = 0; j < ITEMS_PER_THREAD; j++)
2281-
{
2282-
int col_idx = warp_lane+(j*32);
2283-
if(col_idx < valid_items)
2284-
local_data[j] = A[i+col_idx];
2285-
else
2286-
local_data[j] = 0;
2287-
}
2288-
}
2289-
else
2290-
{
2291-
#pragma unroll ITEMS_PER_THREAD
2292-
for(int j = 0; j < ITEMS_PER_THREAD; j++)
2293-
local_data[j] = 0;
2294-
}
2295-
2296-
if(TRANSPOSE)
2297-
{
2298-
#pragma unroll ITEMS_PER_THREAD
2299-
for(int j = 0; j < ITEMS_PER_THREAD; j++)
2300-
{
2301-
int local_col = (32*j)+warp_lane;
2302-
//int local_row = row;
2303-
// store as 256x32
2304-
smem_data[(local_col*33) + row] = local_data[j];
2305-
}
2306-
}
2307-
else
2308-
{
2309-
// treat smem as 32x256, that is 32 rows and 256 columns
2310-
#pragma unroll ITEMS_PER_THREAD
2311-
for(int j = 0; j < ITEMS_PER_THREAD; j++)
2312-
smem_data[row*32*ITEMS_PER_THREAD + (warp_lane) + (j*32)] = local_data[j];
2313-
}
2314-
2315-
2316-
2317-
smem_row += warps;
2318-
2319-
// 1. transpose / reorder in shared memory
2320-
if(smem_row % 32 == 0)
2321-
{
2322-
smem_row = 0;
2323-
__syncthreads();
2324-
2325-
for(int subrow = warp_id; subrow < 32; subrow+=warps)
2326-
{
2327-
for(int j = 0; j < ITEMS_PER_THREAD; j++)
2328-
{
2329-
2330-
switch(FORMAT)
2331-
{
2332-
case COL32:
2333-
if(TRANSPOSE)
2334-
{
2335-
// data lies in shared memory in the following way:
2336-
// row0 [col0 col1 ... col31]
2337-
// row1 [col0 col1 ... col31]
2338-
// ...
2339-
//
2340-
// As such we read consecutive entries with 256 threads (8rows x 32 columns)
2341-
// as j increase, the row increase by a factor of 8
2342-
// We load 8 rows per subrow loop, and subrow increase by 8 per loop
2343-
// so we have an offset of 8 rows every loop or (subrow/warps)*8 = (subrow/8)*8
2344-
const int jrow = j*ITEMS_PER_THREAD; // 8 rows per j
2345-
const int subrow_loop_row = (subrow/warps)*ITEMS_PER_THREAD*ITEMS_PER_THREAD; // 8 rows per j; 8j per subrow loop (subrow/warps)
2346-
//const int local_row = warp_id; // each warp_id is one row
2347-
//const int block_row = base_col; // block offset for row
2348-
//const int local_col = warp_lane
2349-
//const int global_col = base_row; // block offset for col
2350-
if((base_col + subrow_loop_row + jrow + warp_id < outRows) && (base_row+warp_lane < rows))
2351-
{
2352-
// each row has 32 columns and is offset by 1 to prevent bank conflict during storage into smem
2353-
char data = smem_data[(subrow_loop_row + jrow + warp_id)*33 + warp_lane];
2354-
2355-
// each 32 columns we have new tile
2356-
// each tile has size outRows*32 and base_row is done in increments of 32
2357-
offset = base_row*outRows;
2358-
out[offset + (base_col + jrow + subrow_loop_row)*32 + threadIdx.x] = data;
2359-
}
2360-
}
2361-
else
2362-
{
2363-
if(((base_row+subrow) < rows) && (base_col+(j*32)+warp_lane < outCols))
2364-
{
2365-
offset = (base_col/32)*(32*rows);
2366-
char data = smem_data[(subrow*32*ITEMS_PER_THREAD) + (j*32) + warp_lane];
2367-
out[offset+(base_row+subrow)*32 + ((j)*rows*32)+warp_lane] = data;
2368-
}
2369-
}
2370-
break;
2371-
case COL_TURING:
2372-
// TURING FORMAT:
2373-
// 8*32 tiles with 4*4 subtiles
2374-
// the 8*32 subtile has first all 4*4 subtiles of even rows (max 4*4*4 = 64 elements)
2375-
// the subsequent 4*4 subtiles are for all odd rows if some rows columns are empty the values are zero
2376-
// the tile repeats again after the 8*32 tile in a major column order, meaning: (next 8 rows are A[8:16, 0:32])
2377-
// the next tile is the next 8 rows for the same 32 columns. Once all rows are finished, the column
2378-
// index increases by 32
2379-
//
2380-
// [0 0 0 0, 2 2 2 2, 4 4 4 4, 6 6 6 6, 0 0 0 0 ...]
2381-
if(TRANSPOSE)
2382-
{
2383-
const int jrow = j*ITEMS_PER_THREAD; // 8 rows per j
2384-
const int subrow_loop_row = (subrow/warps)*ITEMS_PER_THREAD*ITEMS_PER_THREAD; // 8 rows per j; 8j per subrow loop (subrow/warps)
2385-
//const int local_row = warp_id; // each warp_id is one row
2386-
//const int block_row = base_col; // block offset for row
2387-
//const int local_col = warp_lane
2388-
//const int global_col = base_row; // block offset for col
2389-
if((base_col + subrow_loop_row + jrow + warp_id < outRows) && (base_row+warp_lane < rows))
2390-
{
2391-
// each row has 32 columns and is offset by 1 to prevent bank conflict during storage into smem
2392-
char data = smem_data[(subrow_loop_row + jrow + warp_id)*33 + warp_lane];
2393-
2394-
// each 32 columns we have new tile
2395-
// each tile has size 8*32 = 256 elements offset
2396-
// for each row offset of 8 we increaes the tile first
2397-
// after all rows are exhausted, we increase the col
2398-
int row_offset = ((base_col+jrow+subrow_loop_row+warp_id)/8)*256; // global_row+jrow+subrow_loop_row+local_row, increase tile(=256) every 8 rows
2399-
2400-
// we increase by row_tile_column every 32 columns
2401-
// base_row increase in increments of 32
2402-
//int row_tile_column = 256*outRows/8; // there are outRows/8 row tiles, and each tile is 256 elements
2403-
//int col_offset = (base_row/32)*row_tile_column;
2404-
// -> we can remove the divisions to speed up compute since outRows is always a multiple of 8
2405-
// 256*outRows/8*base_row/32 = outRows*base_row
2406-
int col_offset = outRows*base_row;
2407-
2408-
offset = row_offset+col_offset;
2409-
2410-
// since we process even number of rows with each j (8) and with each subrow (8j) we can determine
2411-
// odd or even rows with the warp_id (each warp processes one row)
2412-
// the col is warp_lane (max 32 columns per row) and the row warp_id
2413-
if(warp_id % 2 == 1)
2414-
// odd
2415-
offset += 128 + (warp_lane/4)*16 + (warp_lane%4) + (((warp_id%8)-1)*2);
2416-
else
2417-
// even
2418-
offset += 0 + (warp_lane/4)*16 + (warp_lane%4) + ((warp_id%8)*2);
2419-
2420-
out[offset] = data;
2421-
}
2422-
}
2423-
else
2424-
{
2425-
if(((base_row+subrow) < rows) && (base_col+(j*32)+warp_lane < outCols))
2426-
{
2427-
char data = smem_data[(subrow*32*ITEMS_PER_THREAD) + (j*32) + warp_lane];
2428-
// set offset designates the tile offset among the 8*32 tiles
2429-
// we first increase rows and then columns. Since we load 128 columns at once
2430-
// we increase the offset by outRows*32 every 32 columns
2431-
// additionally, we increase the offset by 8*32=256 every 8 rows
2432-
offset = ((base_col+(j*32))/32)*outRows*32 + (((base_row+subrow)/8)*256); // global offset (8x32 tile)
2433-
// first 4 rows are reserved for even rows, [0, 2, 4, 6], the next 4 for odd
2434-
// each of these has 32 values in total for 32*4 = 128 as offset if odd
2435-
// every set of 4 columns increases the total offset by 16
2436-
// each even row increase the offset by 4, for example row 2 is offset by 4, 4 by 6 etc so: subrow/2*4 = subrow*2
2437-
// this happens every 8 rows anew (subrow % 8)
2438-
// one writes 4 columns at once that is (col % 4) for the particular index in the subtile
2439-
int subcol = warp_lane;
2440-
2441-
// add local offset (4x4 sub-tile)
2442-
if(subrow % 2 == 1)
2443-
// odd
2444-
offset += 128 + (subcol/4)*16 + (subcol%4) + (((subrow%8)-1)*2);
2445-
else
2446-
// even
2447-
offset += 0 + (subcol/4)*16 + (subcol%4) + ((subrow%8)*2);
2448-
2449-
out[offset] = data;
2450-
}
2451-
}
2452-
break;
2453-
case COL_AMPERE:
2454-
// AMPERE FORMAT:
2455-
// 32*32 tiles with 8*32 subtiles. The rows are interleaved in pairs of two rows with offset of 8 between pairs of two rows:
2456-
// row idx (each number stands for 32 values): [0 1 8 9 16 17 24 25] [2 3 10 11 18 19 26 27]...
2457-
// the tiles are column-major ordered, so after 1024*1024 values we process: A[32:64, 0:32]
2458-
if(TRANSPOSE)
2459-
{
2460-
const int jrow = j*ITEMS_PER_THREAD; // 8 rows per j
2461-
const int subrow_loop_row = (subrow/warps)*ITEMS_PER_THREAD*ITEMS_PER_THREAD; // 8 rows per j; 8j per subrow loop (subrow/warps)
2462-
//const int local_row = warp_id; // each warp_id is one row
2463-
//const int block_row = base_col; // block offset for row
2464-
//const int local_col = warp_lane
2465-
//const int global_col = base_row; // block offset for col
2466-
if((base_col + subrow_loop_row + jrow + warp_id < outRows) && (base_row+warp_lane < rows))
2467-
{
2468-
// each row has 32 columns and is offset by 1 to prevent bank conflict during storage into smem
2469-
char data = smem_data[(subrow_loop_row + jrow + warp_id)*33 + warp_lane];
2470-
2471-
// each 32 columns we have new tile
2472-
// each tile has size 32*32 = 1024 elements offset
2473-
// for each row offset of 32 we increaes the tile first
2474-
// after all rows are exhausted, we increase the col
2475-
int row_offset = ((base_col+jrow+subrow_loop_row+warp_id)/32)*1024; // global_row+jrow+subrow_loop_row+local_row, increase tile(=256) every 8 rows
2476-
2477-
// we increase by row_tile_column every 32 columns
2478-
// base_row increase in increments of 32
2479-
//int row_tile_column = 1024*outRows/32; // there are outRows/32 row tiles, and each tile is 1024 elements
2480-
//int col_offset = (base_row/32)*row_tile_column;
2481-
// -> we can remove the divisions to speed up compute since outRows is always a multiple of 8
2482-
// 1024*outRows/32*base_row/32 = outRows*base_row
2483-
int col_offset = outRows*base_row;
2484-
2485-
offset = row_offset+col_offset;
2486-
2487-
2488-
// same as in the non-transpose case (see below)
2489-
// the difference is that now rows = cols
2490-
// in this case warp_id = subrow
2491-
2492-
// [0 1 8 9 16 17 24 25] [2 3 10 11 18 19 26 27]...
2493-
// subrow % 8 -> [0,1] in tile0, [2, 3] in tile 1 etc
2494-
// subrow % 2 -> 0 for 1st row in the pair, 1 for the 2nd row
2495-
// every 2 rows, the offset increases by two [0, 1, 8, 9...]
2496-
// every 2 rows, the row index increase by 8 [0, 1, 8, 9...]
2497-
int local_row = (jrow + warp_id) % 32; // offset for row > 32 is already calculated into row_offset
2498-
int ampere_row = ((local_row % 8)/2)*8 + (local_row/8)*2 + (local_row % 2);
2499-
2500-
// global offset + row with 32 cols each + 32 cols per j + col_idx=warp_lane
2501-
out[offset + (ampere_row*32) + warp_lane] = data;
2502-
}
2503-
}
2504-
else
2505-
{
2506-
if(((base_row+subrow) < rows) && (base_col+(j*32)+warp_lane < outCols))
2507-
{
2508-
char data = smem_data[(subrow*32*ITEMS_PER_THREAD) + (j*32) + warp_lane];
2509-
2510-
// set offset designates the tile offset among the 32*32 tiles
2511-
// we first increase rows and then columns. Since we load 128 columns at once
2512-
// we increase the offset by outRows*32 every 32 columns
2513-
// additionally, we increase the offset by 32*32=1024 every 32 rows
2514-
offset = ((base_col+(j*32))/32)*outRows*32 + (((base_row+subrow)/32)*1024); // global offset (32x32 tile)
2515-
2516-
// [0 1 8 9 16 17 24 25] [2 3 10 11 18 19 26 27]...
2517-
// subrow % 8 -> [0,1] in tile0, [2, 3] in tile 1 etc
2518-
// subrow % 2 -> 0 for 1st row in the pair, 1 for the 2nd row
2519-
// every 2 rows, the offset increases by two [0, 1, 8, 9...]
2520-
// every 2 rows, the row index increase by 8 [0, 1, 8, 9...]
2521-
int local_row = ((subrow % 8)/2)*8 + (subrow/8)*2 + (subrow % 2);
2522-
2523-
// global offset + row with 32 cols each + 32 cols per j + col_idx
2524-
out[offset + (local_row*32) + warp_lane] = data;
2525-
}
2526-
}
2527-
break;
2528-
}
2529-
}
2530-
}
2531-
}
2532-
}
2533-
}
2534-
25352208
#define DENORM 1.0f/127.0f
25362209
#define MAX_SPARSE_COUNT 32
25372210
#define SMEM_SIZE 8*256
@@ -3386,13 +3059,6 @@ template __global__ void kspmm_coo_very_sparse_naive<signed char, 8, 8>(int *max
33863059
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);
33873060
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);
33883061

3389-
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);
3390-
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);
3391-
template __global__ void kTransformRowToFormat<256, 8, 32, 32*8, 0, COL_TURING>(char *__restrict__ const A, char *out, int rows, int cols, int tiledCols, int outRows, int outCols);
3392-
template __global__ void kTransformRowToFormat<256, 8, 32, 32*8, 1, COL_TURING>(char *__restrict__ const A, char *out, int rows, int cols, int tiledCols, int outRows, int outCols);
3393-
template __global__ void kTransformRowToFormat<256, 8, 32, 32*8, 0, COL_AMPERE>(char *__restrict__ const A, char *out, int rows, int cols, int tiledCols, int outRows, int outCols);
3394-
template __global__ void kTransformRowToFormat<256, 8, 32, 32*8, 1, COL_AMPERE>(char *__restrict__ const A, char *out, int rows, int cols, int tiledCols, int outRows, int outCols);
3395-
33963062
template __global__ void kdequant_mm_int32_fp16<4, 512>(int *__restrict__ const A, float *__restrict__ const rowStats, float *__restrict__ const colStats, half *out, half * __restrict__ const bias, const int numRows, const int numCols, const int n);
33973063

33983064
template __device__ unsigned char dQuantize<0>(float* smem_code, const float rand, float x);

0 commit comments

Comments
 (0)