Skip to content

GitLab

Commit 5b1e2442 authored by Harisankar Sadasivan's avatar Harisankar Sadasivan
Browse files

2-tile sk+ DP with atomics for FP16

parent 2ae16e90
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 697 additions and 722 deletions
example/01_gemm/README.md
View file @ 5b1e2442
......@@ -21,3 +21,25 @@ Warm up
Start running 5 times...
Perf: 1.19685 ms, 107.657 TFlops, 78.8501 GB/s
```
# Instructions for ```example_gemm_xdl_streamk```
## Run ```example_gemm_xdl_streamk```
```bash
# arg1: verification (0=no, 1=yes)
# arg2: initialization (0=no init, 1=integer value, 2=decimal value)
# arg3: time kernel (0=no, 1=yes)
# arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC
# arg10: NumSKBlocks(optional, defaults to DP GEMM)
bin/example_gemm_xdl_streamk 1 2 1 3840 4096 4096 4096 4096 4096 312
```
Result (MI250 @ 1700Mhz, 181TFlops peak FP16 on 1 dye)
```
a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
b_k_n: dim 2, lengths {4096, 4096}, strides {4096, 1}
c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
Recommended grid size :312
Perf: 1.21689 ms, 105.884 TFlops, 79.2748 GB/s, GemmXdlStreamK_RRR_B256_Vec8x2x8_128x128x4x8
```
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_streamk.hpp
View file @ 5b1e2442
......@@ -137,6 +137,8 @@ struct DeviceGemmXdlStreamK : public DeviceGemmStreamK<ALayout,
"setting");
}
// stream-k: calculate the number of blocks to be launched based on #CUs and #occupancy
// dim3 grid_dims = karg.block_mapping.get_grid_dims(karg.num_cu, karg.occupancy);
dim3 grid_dims = karg.block_mapping.get_grid_dims();
float ave_time = 0;
......@@ -268,22 +270,19 @@ struct DeviceGemmXdlStreamK : public DeviceGemmStreamK<ALayout,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
uint32_t NumSKBlocks = 0xffffffff)
uint32_t NumSKBlocks = 0)
{
const auto kernel = kernel_gemm_xdlops_streamk<GridwiseGemm>;
int occupancy, num_cu;
hipError_t rtn;
rtn = hipOccupancyMaxActiveBlocksPerMultiprocessor(
&occupancy, kernel, BlockSize, GridwiseGemm::GetSharedMemoryNumberOfByte());
hip_check_error(rtn);
hip_check_error(
hipOccupancyMaxActiveBlocksPerMultiprocessor(&occupancy, kernel, BlockSize, 0));
hipDeviceProp_t dev_prop;
hipDevice_t dev;
rtn = hipGetDevice(&dev);
hip_check_error(rtn);
rtn = hipGetDeviceProperties(&dev_prop, dev);
hip_check_error(rtn);
hip_check_error(hipGetDevice(&dev));
hip_check_error(hipGetDeviceProperties(&dev_prop, dev));
num_cu = dev_prop.multiProcessorCount;
printf("Assuming full GPU availability, recommended stream-k grid size for tuning :%0d\n",
num_cu * occupancy);
return Argument{p_a,
p_b,
......@@ -318,17 +317,12 @@ struct DeviceGemmXdlStreamK : public DeviceGemmStreamK<ALayout,
{
const auto kernel = kernel_gemm_xdlops_streamk<GridwiseGemm>;
int occupancy, num_cu;
hipError_t rtn;
rtn = hipOccupancyMaxActiveBlocksPerMultiprocessor(
&occupancy, kernel, BlockSize, GridwiseGemm::GetSharedMemoryNumberOfByte());
hip_check_error(rtn);
hip_check_error(
hipOccupancyMaxActiveBlocksPerMultiprocessor(&occupancy, kernel, BlockSize, 0));
hipDeviceProp_t dev_prop;
hipDevice_t dev;
rtn = hipGetDevice(&dev);
hip_check_error(rtn);
rtn = hipGetDeviceProperties(&dev_prop, dev);
hip_check_error(rtn);
hip_check_error(hipGetDevice(&dev));
hip_check_error(hipGetDeviceProperties(&dev_prop, dev));
num_cu = dev_prop.multiProcessorCount;
return std::make_unique<Argument>(reinterpret_cast<const ADataType*>(p_a),
......
include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
View file @ 5b1e2442
......@@ -1010,142 +1010,69 @@ struct BlockToCTileMap_GemmStreamK
MDiv eqav_tiles_big; // for reduction
MDiv eqav_tiles_little; // for reduction
// MDiv tile_swizzle_sub_m_rem;
//--------------------------------------
// prefer construct on host
BlockToCTileMap_GemmStreamK(uint32_t m,
uint32_t n,
uint32_t k,
uint32_t num_cu,
uint32_t occupancy,
uint32_t sk_blocks = 0xffffffff)
uint32_t sk_blocks = 0)
{
// total output tiles
uint32_t num_tiles =
math::integer_divide_ceil(m, MPerBlock) * math::integer_divide_ceil(n, NPerBlock);
k_iters_per_tile = MDiv(math::integer_divide_ceil(k, KPerBlock));
// one cu can hold one wg at one time, from the whole chip's point of view
// if number of wg is same as num_cu, we call it 1 dispatch
// if number of wg is 2x num_cu, we call it 2 dispatches.
// one dispatch can deliver wg same as num_cu (full dispatch), or less than num_cu (partial
// dispatch)
//
uint32_t full_dispatches = num_tiles / num_cu;
uint32_t full_dispatch_tiles = full_dispatches * num_cu;
uint32_t partial_dispatche_tiles = num_tiles - full_dispatch_tiles;
uint32_t sk_occupancy = occupancy;
uint32_t dp_tiles = full_dispatch_tiles;
uint32_t sk_tiles = partial_dispatche_tiles;
if(full_dispatches < occupancy)
{
// in this case, we allocate all blocks as sk blocks
// sk_occupancy = occupancy - full_dispatches;
sk_occupancy = 1; // TODO: single occ seems better
dp_tiles = full_dispatch_tiles;
sk_tiles = partial_dispatche_tiles;
}
else if((occupancy > 1) && (full_dispatches % occupancy == occupancy - 1))
uint32_t dp_tiles, dp_num_blocks, sk_total_iters;
// default to regular DP GEMM if sk blocks == 0
sk_num_blocks = sk_blocks;
if(sk_num_blocks == 0 || sk_num_blocks == 0xFFFFFFFF)
{
// e.g. occupancy = 2, full_dispatches = 3, 5, 7 ...
// occupancy = 3, full_dispatches = 5, 8, 11 ...
// occupancy = 4, full_dispatches = 7, 11 ...
sk_occupancy = 1; // left 1 slot for sk occupancy
dp_tiles = full_dispatch_tiles;
sk_tiles = partial_dispatche_tiles;
sk_num_blocks = 0;
dp_tiles = num_tiles;
sk_num_big_blocks = 0;
k_iters_per_big_block = 0;
dp_num_blocks = num_tiles; // all tile to be dp block
dp_start_block_idx = 0;
sk_total_iters = 0; // clear this tiles
}
// 2-tile sk + DP GEMM
else
{
// others, we reduce 1 dispatch from dp, together with partial dispatch,
// to construct sk dispatch
sk_occupancy = occupancy - ((full_dispatches - 1) % occupancy);
dp_tiles = full_dispatch_tiles - num_cu;
sk_tiles = partial_dispatche_tiles + num_cu;
// grid size
uint32_t grid_size = occupancy * num_cu;
// check if there's enough work for DP+ stream-k
bool bigEnough = num_tiles > grid_size;
// max of 2 sk tiles per block
uint32_t sk_tiles = bigEnough ? grid_size + num_tiles % grid_size : num_tiles;
// remaining tiles are DP tiles
dp_tiles = bigEnough ? (num_tiles - sk_tiles) : 0;
sk_total_iters = k_iters_per_tile.get() * sk_tiles;
// k_iters_per_sk_block is the floor of avg each ck block loop over tiles.
// we need to decide how many iters for each sk block
// let m = k_iters_per_sk_block
// some of the sk block (little) will cover m iters, some (big) will cover m+1
// we have
// 1) l + b = sk_blocks
// 2) l * m + b * (m + 1) = sk_total_iters
// => (l + b) * m + b = sk_total_iters
// => sk_blocks * m + b = sk_total_iters
// => b = sk_total_iters - m * sk_blocks
// NOTE: big could be zero
uint32_t k_iters_per_sk_block = sk_total_iters / sk_num_blocks;
sk_num_big_blocks = sk_total_iters - k_iters_per_sk_block * sk_num_blocks;
k_iters_per_big_block = k_iters_per_sk_block + 1;
dp_num_blocks = dp_tiles;
dp_start_block_idx = (sk_num_blocks + num_cu - 1) / num_cu * num_cu;
}
// uint32_t dp_iters_per_block = k_iters_per_tile.get();
uint32_t sk_total_iters = k_iters_per_tile.get() * sk_tiles;
uint32_t dp_num_blocks = 0;
{
uint32_t min_sk_tiles = (sk_tiles >= num_cu) ? num_cu : (sk_tiles + 1);
uint32_t max_sk_tiles =
(sk_tiles >= num_cu) ? num_cu * sk_occupancy
: math::min(num_cu, sk_total_iters / min_k_iters_per_sk_block);
// if use dp for sk-block, how many iters do we need
uint32_t dp_for_sk_iters = k_iters_per_tile.get();
uint32_t best_sk_score =
std::numeric_limits<int>::max(); // we need to find the smallest sk iters
for(uint32_t tentative_sk_blocks = min_sk_tiles; tentative_sk_blocks < max_sk_tiles;
tentative_sk_blocks++)
{
uint32_t tentative_sk_iters_per_block =
(sk_total_iters + tentative_sk_blocks - 1) / tentative_sk_blocks;
uint32_t tentative_sk_iters = tentative_sk_iters_per_block;
uint32_t sk_blocks_per_tile = (tentative_sk_blocks + sk_tiles - 1) / sk_tiles;
// TODO: carefully adjust this parameter
// the more sk_blocks_per_tile, the worse the overhead
uint32_t cross_sk_blocks_overhead = sk_blocks_per_tile;
if(tentative_sk_blocks % sk_tiles != 0)
{
// penalty for uneven divide
cross_sk_blocks_overhead +=
sk_blocks_per_tile * tentative_sk_iters_per_block / 50;
}
uint32_t tentative_sk_score = tentative_sk_iters + cross_sk_blocks_overhead;
if(tentative_sk_score < best_sk_score)
{
best_sk_score = tentative_sk_score;
sk_num_blocks = tentative_sk_blocks;
}
}
if(best_sk_score >= dp_for_sk_iters)
{
sk_num_blocks = 0;
}
// give a chance to control num of sk blocks
sk_num_blocks = sk_blocks != 0xffffffff ? sk_blocks : sk_num_blocks;
if(sk_num_blocks == 0)
{
sk_num_big_blocks = 0;
k_iters_per_big_block = 0;
dp_num_blocks = num_tiles; // all tile to be dp block
dp_start_block_idx = 0;
sk_total_iters = 0; // clear this tiles
}
else
{
// k_iters_per_sk_block is the floor of avg each ck block loop over tiles.
// we need to decide how many iters for each sk block
// let m = k_iters_per_sk_block
// some of the sk block (little) will cover m iters, some (big) will cover m+1
// we have
// 1) l + b = sk_blocks
// 2) l * m + b * (m + 1) = sk_total_iters
// => (l + b) * m + b = sk_total_iters
// => sk_blocks * m + b = sk_total_iters
// => b = sk_total_iters - m * sk_blocks
// NOTE: big could be zero
uint32_t k_iters_per_sk_block = sk_total_iters / sk_num_blocks;
sk_num_big_blocks = sk_total_iters - k_iters_per_sk_block * sk_num_blocks;
k_iters_per_big_block = k_iters_per_sk_block + 1;
dp_num_blocks = dp_tiles;
dp_start_block_idx = (sk_num_blocks + num_cu - 1) / num_cu * num_cu;
}
}
n_tiles = MDiv2(math::integer_divide_ceil(n, NPerBlock));
n_tiles = MDiv2(math::integer_divide_ceil(n, NPerBlock));
// using multiple blocks for parallel reduction
reduction_start_block_idx = dp_start_block_idx + dp_num_blocks;
if constexpr(ReductionStrategy == StreamKReductionStrategy::Reduction)
......@@ -1157,13 +1084,14 @@ struct BlockToCTileMap_GemmStreamK
}
#if 0
printf("cu:%d, occupancy:%d, grids:%d, num_tiles:%d, dp_tiles:%d, sk_num_big_blocks:%d, "
printf("cu:%d, occupancy:%d, gridsize:%d, num_tiles:%d, dp_tiles:%d, sk_num_big_blocks:%d, "
"sk_num_blocks:%d, "
"sk_total_iters:%d, dp_start_block_idx:%d, dp_iters_per_block:%d, dp_num_blocks:%d, "
"sk_total_iters:%d, dp_start_block_idx:%d, dp_num_blocks:%d, "
"k_iters_per_tile:%d, k_iters_per_big_block:%d, reduction_start_block_idx:%u, "
"sk_tiles:%u, workspace(acc float):%u\n",
num_cu,
occupancy,
// get_grid_dims(num_cu, occupancy).x,
get_grid_dims().x,
num_tiles,
dp_tiles,
......@@ -1171,7 +1099,7 @@ struct BlockToCTileMap_GemmStreamK
sk_num_blocks,
sk_total_iters,
dp_start_block_idx,
dp_iters_per_block,
dp_num_blocks,
k_iters_per_tile.get(),
k_iters_per_big_block,
......@@ -1195,7 +1123,8 @@ struct BlockToCTileMap_GemmStreamK
return k_iters_per_tile.div(sk_total_iters);
}
__host__ __device__ dim3 get_grid_dims() const
// __host__ __device__ constexpr dim3 get_grid_dims(int num_cu, int occupancy) const
__host__ __device__ constexpr dim3 get_grid_dims() const
{
if constexpr(ReductionStrategy == StreamKReductionStrategy::Reduction)
{
......@@ -1203,6 +1132,16 @@ struct BlockToCTileMap_GemmStreamK
}
else
return dim3(reduction_start_block_idx, 1, 1);
// return dim3(num_cu * occupancy, 1, 1); // HS
}
__host__ __device__ uint32_t total_blocks_allocated() const
{
if constexpr(ReductionStrategy == StreamKReductionStrategy::Reduction)
{
return __builtin_amdgcn_readfirstlane(reduction_start_block_idx + get_sk_tiles());
}
else
return __builtin_amdgcn_readfirstlane(reduction_start_block_idx);
}
__device__ uint32_t get_block_idx() const
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment