Skip to content

GitLab

"lightx2v/git@developer.sourcefind.cn:Wenxuan/LightX2V.git" did not exist on "7fde70631ce0b7f67cb2476b44934cde93a2944d"
Unverified Commit 7bcaf2a7 authored by Adam Osewski's avatar Adam Osewski Committed by GitHub
Browse files

Merge branch 'develop' into wavelet_model

parents e59daa22 0345963e
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 2006 additions and 890 deletions
include/ck/tensor_operation/gpu/element/quantization_operation.hpp
View file @ 7bcaf2a7
......@@ -10,8 +10,8 @@ namespace element_wise {
template <typename Activation>
struct Activation_Mul_Clamp
{
Activation_Mul_Clamp(float multiplier, Activation activationOp)
: multiplier_(multiplier), activationOp_(activationOp)
Activation_Mul_Clamp(float requantScale, Activation activationOp)
: requantScale_(requantScale), activationOp_(activationOp)
{
}
......@@ -19,7 +19,7 @@ struct Activation_Mul_Clamp
{
float x_fp32 = ck::type_convert<float>(x);
activationOp_(x_fp32, x_fp32);
float y_fp32 = math::clamp(multiplier_ * x_fp32, -128.f, 127.f);
float y_fp32 = math::clamp(requantScale_ * x_fp32, -128.f, 127.f);
y = ck::type_convert<int8_t>(y_fp32);
}
......@@ -28,10 +28,29 @@ struct Activation_Mul_Clamp
// We might type_convert to int8 after lambda in someplace
float x_fp32 = ck::type_convert<float>(x);
activationOp_(x_fp32, x_fp32);
y = math::clamp(multiplier_ * x_fp32, -128.f, 127.f);
y = math::clamp(requantScale_ * x_fp32, -128.f, 127.f);
}
float requantScale_;
Activation activationOp_;
};
// Conv Perchannel quantization + Activation function which is piecewise linear function, such as
// relu, leaky relu ...etc
template <typename Activation>
struct Activation_Mul2_Clamp
{
Activation_Mul2_Clamp(Activation activationOp) : activationOp_(activationOp) {}
__host__ __device__ constexpr void
operator()(int8_t& y, const int32_t& x, const float& requantScale) const
{
float y_fp32 = ck::type_convert<float>(x);
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(requantScale * y_fp32, -128.f, 127.f);
y = ck::type_convert<int8_t>(y_fp32);
}
float multiplier_;
Activation activationOp_;
};
......@@ -39,21 +58,40 @@ struct Activation_Mul_Clamp
template <typename Activation>
struct Add_Activation_Mul_Clamp
{
Add_Activation_Mul_Clamp(float multiplier, Activation activationOp)
: multiplier_(multiplier), activationOp_(activationOp)
Add_Activation_Mul_Clamp(float requantScale, Activation activationOp)
: requantScale_(requantScale), activationOp_(activationOp)
{
}
__host__ __device__ constexpr void
operator()(int8_t& y, const int32_t& x1, const int32_t& x2) const
operator()(int8_t& y, const int32_t& x, const int32_t& bias) const
{
float y_fp32 = ck::type_convert<float>(x + bias);
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(requantScale_ * y_fp32, -128.f, 127.f);
y = ck::type_convert<int8_t>(y_fp32);
}
float requantScale_;
Activation activationOp_;
};
// Conv Perchannel quantization + Activation function which is piecewise linear function, such as
// relu, leaky relu ...etc
template <typename Activation>
struct Add_Activation_Mul2_Clamp
{
Add_Activation_Mul2_Clamp(Activation activationOp) : activationOp_(activationOp) {}
__host__ __device__ constexpr void
operator()(int8_t& y, const int32_t& x, const int32_t& bias, const float& requantScale) const
{
float y_fp32 = ck::type_convert<float>(x1 + x2);
float y_fp32 = ck::type_convert<float>(x + bias);
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(multiplier_ * y_fp32, -128.f, 127.f);
y_fp32 = math::clamp(requantScale * y_fp32, -128.f, 127.f);
y = ck::type_convert<int8_t>(y_fp32);
}
float multiplier_;
Activation activationOp_;
};
......@@ -61,23 +99,23 @@ struct Add_Activation_Mul_Clamp
template <typename Activation>
struct Add_Mul_Activation_Mul_Clamp
{
Add_Mul_Activation_Mul_Clamp(float multiplier1, float multiplier2, Activation activationOp)
: multiplier1_(multiplier1), multiplier2_(multiplier2), activationOp_(activationOp)
Add_Mul_Activation_Mul_Clamp(float requantScale1, float requantScale2, Activation activationOp)
: requantScale1_(requantScale1), requantScale2_(requantScale2), activationOp_(activationOp)
{
}
__host__ __device__ constexpr void
operator()(int8_t& y, const int32_t& x1, const int32_t& x2) const
operator()(int8_t& y, const int32_t& x, const int32_t& bias) const
{
float y_fp32 = ck::type_convert<float>(x1 + x2);
y_fp32 = multiplier1_ * y_fp32;
float y_fp32 = ck::type_convert<float>(x + bias);
y_fp32 = requantScale1_ * y_fp32;
activationOp_(y_fp32, y_fp32);
y_fp32 = math::clamp(multiplier2_ * y_fp32, -128.f, 127.f);
y_fp32 = math::clamp(requantScale2_ * y_fp32, -128.f, 127.f);
y = ck::type_convert<int8_t>(y_fp32);
}
float multiplier1_;
float multiplier2_;
float requantScale1_;
float requantScale2_;
Activation activationOp_;
};
......
include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp
View file @ 7bcaf2a7
......@@ -16,7 +16,7 @@ template <typename GridwiseReduceSecondHalfBatchNormBackwardFinal_,
typename DyDataType,
typename DxDataType,
typename ScaleDataType,
typename BiasDataType,
typename DscaleDbiasDataType,
typename MeanVarDataType,
typename DyElementwiseOp,
typename XYGridDesc_M_K,
......@@ -35,8 +35,8 @@ __global__ void kernel_reduce_second_half_batchnorm_backward_final(
long_index_t reduce_size,
index_t num_xy_k_block_tile_iteration,
index_t num_dscale_dbias_k_block_tile_iteration,
const ScaleDataType* const __restrict__ p_reduce_dscale,
const BiasDataType* const __restrict__ p_reduce_dbias,
const DscaleDbiasDataType* const __restrict__ p_reduce_dscale,
const DscaleDbiasDataType* const __restrict__ p_reduce_dbias,
const MeanVarDataType* const __restrict__ p_mean,
const MeanVarDataType* const __restrict__ p_inv_var,
const XDataType* const __restrict__ p_x,
......@@ -44,8 +44,8 @@ __global__ void kernel_reduce_second_half_batchnorm_backward_final(
const ScaleDataType* const __restrict__ p_scale,
const DyElementwiseOp dy_elementwise_op,
DxDataType* const __restrict__ p_dx,
ScaleDataType* const __restrict__ p_dscale,
BiasDataType* const __restrict__ p_dbias)
DscaleDbiasDataType* const __restrict__ p_dscale,
DscaleDbiasDataType* const __restrict__ p_dbias)
{
GridwiseReduceSecondHalfBatchNormBackwardFinal_::Run(x_grid_desc_m_k,
dy_grid_desc_m_k,
......@@ -76,7 +76,7 @@ template <typename XDataType,
typename DxDataType,
typename AccDataType,
typename ScaleDataType,
typename BiasDataType,
typename DscaleDbiasDataType,
typename MeanVarDataType,
typename DyElementwiseOp,
typename XYGridDesc_M_K,
......@@ -92,8 +92,8 @@ template <typename XDataType,
index_t XSrcVectorSize,
index_t DySrcVectorSize,
index_t DxDstVectorSize,
index_t ScaleSrcDstVectorSize,
index_t BiasDstVectorSize,
index_t ScaleSrcVectorSize,
index_t DscaleDbiasDstVectorSize,
index_t MeanVarSrcVectorSize>
struct GridwiseReduceSecondHalfBatchNormBackwardFinal
{
......@@ -155,13 +155,13 @@ struct GridwiseReduceSecondHalfBatchNormBackwardFinal
const DscaleDbiasGridDesc_M_K& dscale_dbias_grid_desc_m_k,
const MeanVarGridDesc_M& mean_var_grid_desc_m,
const ScaleBiasGridDesc_M& scale_grid_desc_m,
const ScaleBiasGridDesc_M& bias_grid_desc_m,
const ScaleBiasGridDesc_M& dscale_dbias_grid_desc_m,
index_t blkgroup_size,
long_index_t reduce_size,
index_t num_xy_k_block_tile_iteration,
index_t num_dscale_dbias_k_block_tile_iteration,
const ScaleDataType* const __restrict__ p_reduce_dscale,
const BiasDataType* const __restrict__ p_reduce_dbias,
const DscaleDbiasDataType* const __restrict__ p_reduce_dscale,
const DscaleDbiasDataType* const __restrict__ p_reduce_dbias,
const MeanVarDataType* const __restrict__ p_mean,
const MeanVarDataType* const __restrict__ p_inv_var,
const XDataType* const __restrict__ p_x,
......@@ -169,8 +169,8 @@ struct GridwiseReduceSecondHalfBatchNormBackwardFinal
const ScaleDataType* const __restrict__ p_scale,
const DyElementwiseOp dy_elementwise_op,
DxDataType* const __restrict__ p_dx,
ScaleDataType* const __restrict__ p_dscale,
BiasDataType* const __restrict__ p_dbias)
DscaleDbiasDataType* const __restrict__ p_dscale,
DscaleDbiasDataType* const __restrict__ p_dbias)
{
__shared__ AccDataType p_reduce_work_buffer[BlockSize];
......@@ -222,24 +222,8 @@ struct GridwiseReduceSecondHalfBatchNormBackwardFinal
// Step 1: do final reduction of dbias = sum(dy), dscale = sum(dy * (x-mean) * inv-variance)
// clang-format on
auto threadwise_dscale_load_m_k =
ThreadwiseTensorSliceTransfer_v2<ScaleDataType,
AccDataType,
DscaleDbiasGridDesc_M_K,
decltype(thread_buffer_desc_m_1),
ThreadBufferLengths_M_1,
Sequence<0, 1>,
1,
1,
1,
true>(
dscale_dbias_grid_desc_m_k,
make_multi_index(blkgroup_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * 1));
auto threadwise_dbias_load_m_k =
ThreadwiseTensorSliceTransfer_v2<BiasDataType,
auto threadwise_dscale_dbias_load_m_k =
ThreadwiseTensorSliceTransfer_v2<DscaleDbiasDataType,
AccDataType,
DscaleDbiasGridDesc_M_K,
decltype(thread_buffer_desc_m_1),
......@@ -254,38 +238,20 @@ struct GridwiseReduceSecondHalfBatchNormBackwardFinal
thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * 1));
auto threadwise_dscale_store_m =
auto threadwise_dscale_dbias_store_m =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
ScaleDataType,
DscaleDbiasDataType,
decltype(thread_buffer_desc_m),
ScaleBiasGridDesc_M,
PassThroughOp,
ThreadBufferLengths_M,
Sequence<0>,
0,
ScaleSrcDstVectorSize,
DscaleDbiasDstVectorSize,
InMemoryDataOperationEnum::Set,
1,
true>(
scale_grid_desc_m,
make_multi_index(blkgroup_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize),
PassThroughOp{});
auto threadwise_dbias_store_m =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
BiasDataType,
decltype(thread_buffer_desc_m),
ScaleBiasGridDesc_M,
PassThroughOp,
ThreadBufferLengths_M,
Sequence<0>,
0,
BiasDstVectorSize,
InMemoryDataOperationEnum::Set,
1,
true>(
bias_grid_desc_m,
dscale_dbias_grid_desc_m,
make_multi_index(blkgroup_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize),
PassThroughOp{});
......@@ -297,10 +263,10 @@ struct GridwiseReduceSecondHalfBatchNormBackwardFinal
p_reduce_dbias, dscale_dbias_grid_desc_m_k.GetElementSpaceSize());
auto dscale_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_dscale, scale_grid_desc_m.GetElementSpaceSize());
p_dscale, dscale_dbias_grid_desc_m.GetElementSpaceSize());
auto dbias_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_dbias, bias_grid_desc_m.GetElementSpaceSize());
p_dbias, dscale_dbias_grid_desc_m.GetElementSpaceSize());
constexpr auto dscale_dbias_thread_copy_step_m_k =
make_multi_index(0, KThreadClusterSize * 1);
......@@ -313,25 +279,23 @@ struct GridwiseReduceSecondHalfBatchNormBackwardFinal
for(index_t reducedTiles = 0; reducedTiles < num_dscale_dbias_k_block_tile_iteration;
++reducedTiles)
{
threadwise_dscale_load_m_k.Run(dscale_dbias_grid_desc_m_k,
reduce_dscale_global_buf,
thread_buffer_desc_m_1,
make_tuple(I0, I0),
reduce_dscale_thread_buf);
threadwise_dbias_load_m_k.Run(dscale_dbias_grid_desc_m_k,
reduce_dbias_global_buf,
thread_buffer_desc_m_1,
make_tuple(I0, I0),
reduce_dbias_thread_buf);
threadwise_dscale_dbias_load_m_k.Run(dscale_dbias_grid_desc_m_k,
reduce_dscale_global_buf,
thread_buffer_desc_m_1,
make_tuple(I0, I0),
reduce_dscale_thread_buf);
threadwise_dscale_dbias_load_m_k.Run(dscale_dbias_grid_desc_m_k,
reduce_dbias_global_buf,
thread_buffer_desc_m_1,
make_tuple(I0, I0),
reduce_dbias_thread_buf);
ThreadwiseReduce::Reduce(reduce_dscale_thread_buf, dscale_thread_buf);
ThreadwiseReduce::Reduce(reduce_dbias_thread_buf, dbias_thread_buf);
threadwise_dscale_load_m_k.MoveSrcSliceWindow(dscale_dbias_grid_desc_m_k,
dscale_dbias_thread_copy_step_m_k);
threadwise_dbias_load_m_k.MoveSrcSliceWindow(dscale_dbias_grid_desc_m_k,
dscale_dbias_thread_copy_step_m_k);
threadwise_dscale_dbias_load_m_k.MoveSrcSliceWindow(dscale_dbias_grid_desc_m_k,
dscale_dbias_thread_copy_step_m_k);
}
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
......@@ -343,17 +307,17 @@ struct GridwiseReduceSecondHalfBatchNormBackwardFinal
BlockwiseReduce::Reduce(reduce_work_buf, dbias_thread_buf(I));
});
threadwise_dscale_store_m.Run(thread_buffer_desc_m,
make_tuple(I0),
dscale_thread_buf,
scale_grid_desc_m,
dscale_global_buf);
threadwise_dscale_dbias_store_m.Run(thread_buffer_desc_m,
make_tuple(I0),
dscale_thread_buf,
dscale_dbias_grid_desc_m,
dscale_global_buf);
threadwise_dbias_store_m.Run(thread_buffer_desc_m,
make_tuple(I0),
dbias_thread_buf,
bias_grid_desc_m,
dbias_global_buf);
threadwise_dscale_dbias_store_m.Run(thread_buffer_desc_m,
make_tuple(I0),
dbias_thread_buf,
dscale_dbias_grid_desc_m,
dbias_global_buf);
// clang-format off
// Step 2: calculate dx = 1/N * inv-variance * scale * (N * dy - dbias - dscale * (x - mean) * inv-variance)
......@@ -418,7 +382,7 @@ struct GridwiseReduceSecondHalfBatchNormBackwardFinal
ThreadBufferLengths_M,
Sequence<0>,
0,
ScaleSrcDstVectorSize,
ScaleSrcVectorSize,
1,
true>(
scale_grid_desc_m,
......
include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp
View file @ 7bcaf2a7
......@@ -17,7 +17,7 @@ template <typename GridwiseWelfordSecondHalfReduceFirstHalf_,
typename DyDataType,
typename AccDataType,
typename ScaleDataType,
typename BiasDataType,
typename DscaleDbiasDataType,
typename MeanVarDataType,
typename DyElementwiseOp,
typename XYGridDesc_M_K,
......@@ -45,8 +45,8 @@ __global__ void kernel_welford_second_half_reduce_first_half(
MeanVarDataType* const __restrict__ p_out_welford_inv_variance,
const XDataType* const __restrict__ p_x,
const DyDataType* const __restrict__ p_dy,
ScaleDataType* const __restrict__ p_reduce_dscale,
BiasDataType* const __restrict__ p_reduce_dbias)
DscaleDbiasDataType* const __restrict__ p_reduce_dscale,
DscaleDbiasDataType* const __restrict__ p_reduce_dbias)
{
GridwiseWelfordSecondHalfReduceFirstHalf_::Run(x_grid_desc_m_k,
dy_grid_desc_m_k,
......@@ -76,7 +76,7 @@ template <typename XDataType,
typename DyDataType,
typename AccDataType,
typename ScaleDataType,
typename BiasDataType,
typename DscaleDbiasDataType,
typename MeanVarDataType,
typename DyElementwiseOp,
typename XYGridDesc_M_K,
......@@ -174,8 +174,8 @@ struct GridwiseWelfordSecondHalfReduceFirstHalf
MeanVarDataType* const __restrict__ p_out_welford_inv_variance,
const XDataType* const __restrict__ p_x,
const DyDataType* const __restrict__ p_dy,
ScaleDataType* const __restrict__ p_reduce_dscale,
BiasDataType* const __restrict__ p_reduce_dbias)
DscaleDbiasDataType* const __restrict__ p_reduce_dscale,
DscaleDbiasDataType* const __restrict__ p_reduce_dbias)
{
__shared__ AccDataType p_reduce_work_buffer[BlockSize];
......@@ -511,28 +511,9 @@ struct GridwiseWelfordSecondHalfReduceFirstHalf
BlockwiseReduce::Reduce(reduce_work_buf, reduce_dbias_thread_buf(I));
});
auto threadwise_dscale_store =
auto threadwise_dscale_dbias_store =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
ScaleDataType,
decltype(thread_buffer_desc_m_1),
DscaleDbiasGridDesc_M_G,
PassThroughOp,
ThreadBufferLengths_M_1,
Sequence<0, 1>,
1,
1,
InMemoryDataOperationEnum::Set,
1,
true>(
dscale_dbias_grid_desc_m_g,
make_multi_index(blkgroup_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize,
block_local_id),
PassThroughOp{});
auto threadwise_dbias_store =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
BiasDataType,
DscaleDbiasDataType,
decltype(thread_buffer_desc_m_1),
DscaleDbiasGridDesc_M_G,
PassThroughOp,
......@@ -557,17 +538,17 @@ struct GridwiseWelfordSecondHalfReduceFirstHalf
if(thread_k_cluster_id == 0)
{
threadwise_dscale_store.Run(thread_buffer_desc_m_1,
make_tuple(I0, I0),
reduce_dscale_thread_buf,
dscale_dbias_grid_desc_m_g,
reduce_dscale_global_buf);
threadwise_dbias_store.Run(thread_buffer_desc_m_1,
make_tuple(I0, I0),
reduce_dbias_thread_buf,
dscale_dbias_grid_desc_m_g,
reduce_dbias_global_buf);
threadwise_dscale_dbias_store.Run(thread_buffer_desc_m_1,
make_tuple(I0, I0),
reduce_dscale_thread_buf,
dscale_dbias_grid_desc_m_g,
reduce_dscale_global_buf);
threadwise_dscale_dbias_store.Run(thread_buffer_desc_m_1,
make_tuple(I0, I0),
reduce_dbias_thread_buf,
dscale_dbias_grid_desc_m_g,
reduce_dbias_global_buf);
};
};
};
......
include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp
View file @ 7bcaf2a7
......@@ -796,6 +796,11 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
}
});
}
else
{
static_for<0, acc_thread_buf.Size(), 1>{}(
[&](auto i) { acc_element_op(acc_thread_buf(i), acc_thread_buf[i]); });
}
block_sync_lds(); // wait for lds read in gemm0 blockwise gemm
......
include/ck/utility/amd_wmma.hpp 0 → 100644
View file @ 7bcaf2a7
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#ifndef CK_AMD_WMMA_HPP
#define CK_AMD_WMMA_HPP
#include "data_type.hpp"
// TODO: Add arch limitation
namespace ck {
// wave32 only
// src: fp16, dst: fp32
template <index_t MPerWave, index_t NPerWave>
struct intrin_wmma_f32_16x16x16_f16_w32;
template <>
struct intrin_wmma_f32_16x16x16_f16_w32<16, 16>
{
template <class FloatC>
__device__ static void Run(const half16_t& reg_a, const half16_t& reg_b, FloatC& reg_c)
{
reg_c.template AsType<float8_t>()(Number<0>{}) = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32(
reg_a, reg_b, reg_c.template AsType<float8_t>()[Number<0>{}]);
}
};
// src: bf16, dst: fp32
template <index_t MPerWave, index_t NPerWave>
struct intrin_wmma_f32_16x16x16_bf16_w32;
template <>
struct intrin_wmma_f32_16x16x16_bf16_w32<16, 16>
{
template <class FloatC>
__device__ static void Run(const bhalf16_t& reg_a, const bhalf16_t& reg_b, FloatC& reg_c)
{
reg_c.template AsType<float8_t>()(Number<0>{}) =
__builtin_amdgcn_wmma_f32_16x16x16_bf16_w32(
reg_a, reg_b, reg_c.template AsType<float8_t>()[Number<0>{}]);
}
};
// src: fp16, dst: fp16
template <index_t MPerWave, index_t NPerWave, index_t Opsel>
struct intrin_wmma_f16_16x16x16_f16_w32;
template <index_t Opsel>
struct intrin_wmma_f16_16x16x16_f16_w32<16, 16, Opsel>
{
template <class FloatC>
__device__ static void Run(const half16_t& reg_a, const half16_t& reg_b, FloatC& reg_c)
{
// opsel usage
// false: D0.[0:15] = result
// true : D0.[16:31]= result
reg_c.template AsType<half16_t>()(Number<0>{}) = __builtin_amdgcn_wmma_f16_16x16x16_f16_w32(
reg_a, reg_b, reg_c.template AsType<half16_t>()[Number<0>{}], Opsel);
}
};
// src: bf16, dst: bf16
template <index_t MPerWave, index_t NPerWave, index_t Opsel>
struct intrin_wmma_bf16_16x16x16_bf16_w32;
template <index_t Opsel>
struct intrin_wmma_bf16_16x16x16_bf16_w32<16, 16, Opsel>
{
template <class FloatC>
__device__ static void Run(const bhalf16_t& reg_a, const bhalf16_t& reg_b, FloatC& reg_c)
{
// opsel usage
// false: D0.[0:15] = result
// true : D0.[16:31]= result
reg_c.template AsType<bhalf16_t>()(Number<0>{}) =
__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32(
reg_a, reg_b, reg_c.template AsType<bhalf16_t>()[Number<0>{}], Opsel);
}
};
// src: iu8, dst: i32
template <index_t MPerWave, index_t NPerWave, bool neg_a, bool neg_b, bool clamp>
struct intrin_wmma_i32_16x16x16_iu8_w32;
template <bool neg_a, bool neg_b, bool clamp>
struct intrin_wmma_i32_16x16x16_iu8_w32<16, 16, neg_a, neg_b, clamp>
{
template <class FloatC>
__device__ static void Run(const int8x16_t& reg_a, const int8x16_t& reg_b, FloatC& reg_c)
{
reg_c.template AsType<int32x8_t>()(Number<0>{}) =
__builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
neg_a,
bit_cast<int32x4_t>(reg_a),
neg_b,
bit_cast<int32x4_t>(reg_b),
reg_c.template AsType<int32x8_t>()[Number<0>{}],
clamp);
}
};
} // namespace ck
#endif
include/ck/utility/math_v2.hpp
View file @ 7bcaf2a7
......@@ -114,7 +114,16 @@ static inline __device__ int4_t abs(int4_t x)
};
#endif
static inline __device__ half_t abs(half_t x) { return ::__habs(x); };
static inline __device__ half_t abs(half_t x)
{
uint16_t xx = ck::bit_cast<uint16_t>(x);
uint16_t abs_xx = xx & 0x7fff;
half_t abs_x = ck::bit_cast<half_t>(abs_xx);
return abs_x;
};
static inline __device__ bool isnan(float x) { return ::isnan(x); };
......@@ -140,7 +149,12 @@ static inline __device__ bool isnan(int4_t x)
};
#endif
static inline __device__ bool isnan(half_t x) { return ::__hisnan(x); };
static inline __device__ bool isnan(half_t x)
{
uint16_t xx = ck::bit_cast<uint16_t>(x);
return (xx & 0x7FFF) > 0x7C00;
};
static inline __device__ float sqrt(float x) { return ::sqrtf(x); };
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_bias_forward_perlayer_quantization.hpp → library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp
View file @ 7bcaf2a7
......@@ -23,7 +23,7 @@ void add_device_conv2d_bias_perlayer_quantization_int8_instances(
std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
GKYXC,
GK_TUPLE,
GK_Tuple,
GNHWK,
int8_t,
int8_t,
......@@ -38,7 +38,7 @@ void add_device_conv2d_bias_relu_perlayer_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
GKYXC,
GK_TUPLE,
GK_Tuple,
GNHWK,
int8_t,
int8_t,
......@@ -91,7 +91,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_TUPLE> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_Tuple> &&
is_same_v<OutLayout, GNHWK>)
{
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
......
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