ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

support max3 in smoothquant and add+ rmsnorm + rdquant (#1654)

Browse Source 
* Fix cmake example build

* Support max3 in smoothquant one pass

* support max3 in two pass

* support max3 in add_rmsnorm_rdquant
This commit is contained in:
rocking
2024-11-27 05:01:15 +08:00
committed by GitHub
parent bfe983a151
commit abae2afc72
5 changed files with 94 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
include/ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_one_pass.hpp
View File

@@ -25,6 +25,7 @@ struct SmoothquantPipelineOnePass
static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
static constexpr bool kPadM = false; // TODO - BlockSmoothquantProblem::kPadM
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool UseMax3 = true; // TODO - Move to trait
static constexpr const char* name = []() {
if constexpr(kNeedCrossWarpSync)
@@ -52,7 +53,15 @@ struct SmoothquantPipelineOnePass
xscale_window_, Policy::template MakeXScaleBlockTileDistribution<Problem>());
auto reduce_absmax_func = ReduceOp::AbsMax{};
auto reduce_max_func = ReduceOp::Max{};
auto reduce_absmax3_func = [](auto acc_, auto v_0_, auto v_1_) {
float rtn;
asm volatile("v_max3_f32 %0, %1, abs(%2), abs(%3)"
: "=v"(rtn)
: "v"(acc_), "v"(v_0_), "v"(v_1_));
return rtn;
};
auto reduce_max_func = ReduceOp::Max{};
auto block_reduce2d = Policy::template GetBlockReduce2d<Problem>();
auto block_reduce2d_sync = Policy::template GetBlockReduce2dSync<Problem>();
auto block_reduce2d_cross_warp_sync =
@@ -68,8 +77,23 @@ struct SmoothquantPipelineOnePass
xscale);
// compute absmax, cross-lane->cross-warp
auto absmax = block_reduce2d(
y, reduce_absmax_func.GetIdentityValue<ComputeDataType>(), reduce_absmax_func);
auto absmax = [&]() {
constexpr auto x_size_per_row =
x.get_tile_distribution().get_ys_to_d_descriptor().get_lengths().at(number<1>{});
if constexpr(UseMax3 && std::is_same_v<ComputeDataType, float> &&
x_size_per_row % 2 == 0)
{
return block_reduce2d(y,
reduce_absmax_func.GetIdentityValue<ComputeDataType>(),
reduce_absmax3_func,
sequence<1, 2>{});
}
else
{
return block_reduce2d(
y, reduce_absmax_func.GetIdentityValue<ComputeDataType>(), reduce_absmax_func);
}
}();
block_reduce2d_sync(absmax, reduce_max_func);
block_reduce2d_cross_warp_sync(absmax, smem, reduce_max_func);

16
include/ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_two_pass.hpp
View File

@@ -25,6 +25,7 @@ struct SmoothquantPipelineTwoPass
static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
static constexpr bool kPadM = false; // TODO - BlockSmoothquantProblem::kPadM
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool UseMax3 = true; // TODO - Move to trait
static constexpr const char* name = []() {
if constexpr(kNeedCrossWarpSync)
@@ -56,6 +57,13 @@ struct SmoothquantPipelineTwoPass
__builtin_amdgcn_readfirstlane(integer_divide_ceil(row_size, Block_N));
auto reduce_absmax_func = ReduceOp::AbsMax{};
auto reduce_absmax3_func = [](auto acc_, auto v_0_, auto v_1_) {
float rtn;
asm volatile("v_max3_f32 %0, %1, abs(%2), abs(%3)"
: "=v"(rtn)
: "v"(acc_), "v"(v_0_), "v"(v_1_));
return rtn;
};
auto reduce_max_func = ReduceOp::Max{};
auto block_reduce2d = Policy::template GetBlockReduce2d<Problem>();
auto block_reduce2d_sync = Policy::template GetBlockReduce2dSync<Problem>();
@@ -77,7 +85,13 @@ struct SmoothquantPipelineTwoPass
x,
xscale);
block_reduce2d(y, absmax, reduce_absmax_func);
constexpr auto x_size_per_row =
x.get_tile_distribution().get_ys_to_d_descriptor().get_lengths().at(number<1>{});
if constexpr(UseMax3 && std::is_same_v<ComputeDataType, float> &&
x_size_per_row % 2 == 0)
block_reduce2d(y, absmax, reduce_absmax3_func, sequence<1, 2>{});
else
block_reduce2d(y, absmax, reduce_absmax_func);
move_tile_window(x_window, {0, Block_N});
move_tile_window(xscale_window, {Block_N});