Skip to content

GitLab

Commit 766b0a9e authored by Chao Liu's avatar Chao Liu
Browse files

experimenting

parent f35c64eb
Expand all Show whitespace changes
Inline Side-by-side
Showing with 561 additions and 534 deletions
src/include/functional.hip.hpp
View file @ 766b0a9e
#pragma once
#include "constant_integral.hip.hpp"
template <unsigned NLoop>
template <index_t NLoop>
struct static_loop_n
{
template <class F>
......@@ -24,7 +24,7 @@ struct static_loop_n<1>
}
};
template <unsigned NLoop>
template <index_t NLoop>
struct static_const_reduce_n
{
template <class F, class Reduce>
......
src/include/gridwise_direct_convolution_1.hip.hpp
View file @ 766b0a9e
......@@ -8,18 +8,18 @@ template <class Float,
class InGlobalDesc,
class WeiGlobalDesc,
class OutGlobalDesc,
unsigned NPerBlock,
unsigned KPerBlock,
unsigned CPerBlock,
unsigned HoPerBlock,
unsigned WoPerBlock,
unsigned NPerThread,
unsigned KPerThread,
unsigned CPerThread,
unsigned HoPerThread,
unsigned WoPerThread,
unsigned BlockSize,
unsigned GridSize>
index_t NPerBlock,
index_t KPerBlock,
index_t CPerBlock,
index_t HoPerBlock,
index_t WoPerBlock,
index_t NPerThread,
index_t KPerThread,
index_t CPerThread,
index_t HoPerThread,
index_t WoPerThread,
index_t BlockSize,
index_t GridSize>
__global__ void gridwise_direct_convolution_1(const Float* const __restrict__ p_in_global,
const Float* const __restrict__ p_wei_global,
Float* const __restrict__ p_out_global)
......@@ -33,16 +33,16 @@ __global__ void gridwise_direct_convolution_1(const Float* const __restrict__ p_
constexpr auto wei_global_desc = WeiGlobalDesc{};
constexpr auto out_global_desc = OutGlobalDesc{};
constexpr unsigned Y = wei_global_desc.GetLength(I2);
constexpr unsigned X = wei_global_desc.GetLength(I3);
constexpr index_t Y = wei_global_desc.GetLength(I2);
constexpr index_t X = wei_global_desc.GetLength(I3);
constexpr unsigned HiPerBlock = HoPerBlock + Y - 1;
constexpr unsigned WiPerBlock = WoPerBlock + X - 1;
constexpr index_t HiPerBlock = HoPerBlock + Y - 1;
constexpr index_t WiPerBlock = WoPerBlock + X - 1;
constexpr unsigned NBlockWork = (out_global_desc.GetLength(I0) + NPerBlock - 1) / NPerBlock;
constexpr unsigned KBlockWork = (out_global_desc.GetLength(I1) + KPerBlock - 1) / KPerBlock;
constexpr unsigned HBlockWork = (out_global_desc.GetLength(I2) + HoPerBlock - 1) / HoPerBlock;
constexpr unsigned WBlockWork = (out_global_desc.GetLength(I3) + WoPerBlock - 1) / WoPerBlock;
constexpr index_t NBlockWork = (out_global_desc.GetLength(I0) + NPerBlock - 1) / NPerBlock;
constexpr index_t KBlockWork = (out_global_desc.GetLength(I1) + KPerBlock - 1) / KPerBlock;
constexpr index_t HBlockWork = (out_global_desc.GetLength(I2) + HoPerBlock - 1) / HoPerBlock;
constexpr index_t WBlockWork = (out_global_desc.GetLength(I3) + WoPerBlock - 1) / WoPerBlock;
constexpr auto in_block_global_desc = make_ConstantTensorDescriptor(
Sequence<NPerBlock, CPerBlock, HiPerBlock, WiPerBlock>{}, in_global_desc.GetStrides());
......@@ -59,31 +59,31 @@ __global__ void gridwise_direct_convolution_1(const Float* const __restrict__ p_
constexpr auto out_block_desc =
make_ConstantTensorDescriptor(out_block_global_desc.GetLengths());
constexpr unsigned in_block_size = in_block_desc.GetElementSpace();
constexpr unsigned wei_block_size = wei_block_desc.GetElementSpace();
constexpr unsigned out_block_size = out_block_desc.GetElementSpace();
constexpr index_t in_block_size = in_block_desc.GetElementSpace();
constexpr index_t wei_block_size = wei_block_desc.GetElementSpace();
constexpr index_t out_block_size = out_block_desc.GetElementSpace();
__shared__ Float p_in_block[in_block_size];
__shared__ Float p_wei_block[wei_block_size];
__shared__ Float p_out_block[out_block_size];
const unsigned block_id = blockIdx.x;
const index_t block_id = blockIdx.x;
unsigned itmp = block_id;
unsigned n_block_work_id = itmp / (KBlockWork * HBlockWork * WBlockWork);
index_t itmp = block_id;
index_t n_block_work_id = itmp / (KBlockWork * HBlockWork * WBlockWork);
itmp -= n_block_work_id * (KBlockWork * HBlockWork * WBlockWork);
unsigned k_block_work_id = itmp / (HBlockWork * WBlockWork);
index_t k_block_work_id = itmp / (HBlockWork * WBlockWork);
itmp -= k_block_work_id * (HBlockWork * WBlockWork);
unsigned h_block_work_id = itmp / WBlockWork;
unsigned w_block_work_id = itmp - h_block_work_id * WBlockWork;
index_t h_block_work_id = itmp / WBlockWork;
index_t w_block_work_id = itmp - h_block_work_id * WBlockWork;
unsigned n_block_work_begin = n_block_work_id * NPerBlock;
unsigned k_block_work_begin = k_block_work_id * KPerBlock;
unsigned ho_block_work_begin = h_block_work_id * HoPerBlock;
unsigned wo_block_work_begin = w_block_work_id * WoPerBlock;
index_t n_block_work_begin = n_block_work_id * NPerBlock;
index_t k_block_work_begin = k_block_work_id * KPerBlock;
index_t ho_block_work_begin = h_block_work_id * HoPerBlock;
index_t wo_block_work_begin = w_block_work_id * WoPerBlock;
unsigned hi_block_work_begin = ho_block_work_begin; // minus padding
unsigned wi_block_work_begin = wo_block_work_begin; // minus padding
index_t hi_block_work_begin = ho_block_work_begin; // minus padding
index_t wi_block_work_begin = wo_block_work_begin; // minus padding
constexpr auto blockwise_in_copy =
Blockwise4dTensorCopy1<BlockSize,
......@@ -109,7 +109,7 @@ __global__ void gridwise_direct_convolution_1(const Float* const __restrict__ p_
// set output tensor in LDS to 0
blockwise_4d_tensor_set_zero<BlockSize>(out_block_desc, p_out_block);
for(unsigned c_block_work_begin = 0; c_block_work_begin < in_global_desc.GetLength(I1);
for(index_t c_block_work_begin = 0; c_block_work_begin < in_global_desc.GetLength(I1);
c_block_work_begin += CPerBlock)
{
// copy input tensor to LDS
......
src/include/gridwise_direct_convolution_2_nchw_kcyx_nkhw.hip.hpp
View file @ 766b0a9e
......@@ -11,20 +11,20 @@ template <class Float,
class InGlobalDesc,
class WeiGlobalDesc,
class OutGlobalDesc,
unsigned NPerBlock,
unsigned KPerBlock,
unsigned CPerBlock,
unsigned HoPerBlock,
unsigned WoPerBlock,
unsigned NPerThread,
unsigned KPerThread,
unsigned CPerThread,
unsigned HoPerThread,
unsigned WoPerThread,
unsigned InBlockCopyDataPerRead,
unsigned WeiBlockCopyDataPerRead,
unsigned BlockSize,
unsigned GridSize>
index_t NPerBlock,
index_t KPerBlock,
index_t CPerBlock,
index_t HoPerBlock,
index_t WoPerBlock,
index_t NPerThread,
index_t KPerThread,
index_t CPerThread,
index_t HoPerThread,
index_t WoPerThread,
index_t InBlockCopyDataPerRead,
index_t WeiBlockCopyDataPerRead,
index_t BlockSize,
index_t GridSize>
__global__ void
gridwise_direct_convolution_2_nchw_kcyx_nkhw(const Float* const __restrict__ p_in_global,
const Float* const __restrict__ p_wei_global,
......@@ -39,17 +39,17 @@ gridwise_direct_convolution_2_nchw_kcyx_nkhw(const Float* const __restrict__ p_i
constexpr auto wei_kcyx_global_desc = WeiGlobalDesc{};
constexpr auto out_nkhw_global_desc = OutGlobalDesc{};
constexpr unsigned N = in_nchw_global_desc.GetLength(I0);
constexpr unsigned K = wei_kcyx_global_desc.GetLength(I0);
constexpr unsigned C = wei_kcyx_global_desc.GetLength(I1);
constexpr unsigned Y = wei_kcyx_global_desc.GetLength(I2);
constexpr unsigned X = wei_kcyx_global_desc.GetLength(I3);
constexpr index_t N = in_nchw_global_desc.GetLength(I0);
constexpr index_t K = wei_kcyx_global_desc.GetLength(I0);
constexpr index_t C = wei_kcyx_global_desc.GetLength(I1);
constexpr index_t Y = wei_kcyx_global_desc.GetLength(I2);
constexpr index_t X = wei_kcyx_global_desc.GetLength(I3);
constexpr auto wei_ke_global_desc = make_ConstantTensorDescriptor(
Sequence<K, C * Y * X>{}); // 2d view of wei for blockwise copy
constexpr unsigned HiPerBlock = HoPerBlock + Y - 1;
constexpr unsigned WiPerBlock = WoPerBlock + X - 1;
constexpr index_t HiPerBlock = HoPerBlock + Y - 1;
constexpr index_t WiPerBlock = WoPerBlock + X - 1;
constexpr auto in_nchw_block_desc = make_ConstantTensorDescriptor_aligned(
Sequence<NPerBlock, CPerBlock, HiPerBlock, WiPerBlock>{}, Number<InBlockCopyDataPerRead>{});
......@@ -63,12 +63,12 @@ gridwise_direct_convolution_2_nchw_kcyx_nkhw(const Float* const __restrict__ p_i
Sequence<wei_ke_block_desc.GetStride(I0), Y * X, X, 1>{});
// shared mem
constexpr unsigned in_block_size =
constexpr index_t in_block_size =
in_nchw_block_desc.GetElementSpace(Number<InBlockCopyDataPerRead>{});
constexpr unsigned wei_block_size =
constexpr index_t wei_block_size =
wei_kcyx_block_desc.GetElementSpace(Number<WeiBlockCopyDataPerRead>{});
constexpr unsigned max_align = InBlockCopyDataPerRead > WeiBlockCopyDataPerRead
constexpr index_t max_align = InBlockCopyDataPerRead > WeiBlockCopyDataPerRead
? InBlockCopyDataPerRead
: WeiBlockCopyDataPerRead;
......@@ -76,8 +76,8 @@ gridwise_direct_convolution_2_nchw_kcyx_nkhw(const Float* const __restrict__ p_i
__shared__ Float p_wei_block[max_align * ((wei_block_size + max_align - 1) / max_align)];
// threadwise tensors
constexpr unsigned HiPerThread = HoPerThread + Y - 1;
constexpr unsigned WiPerThread = WoPerThread + X - 1;
constexpr index_t HiPerThread = HoPerThread + Y - 1;
constexpr index_t WiPerThread = WoPerThread + X - 1;
constexpr auto in_nchw_thread_block_desc =
make_ConstantTensorDescriptor(Sequence<NPerThread, CPerThread, HiPerThread, WiPerThread>{},
......@@ -93,56 +93,54 @@ gridwise_direct_convolution_2_nchw_kcyx_nkhw(const Float* const __restrict__ p_i
Float p_out_thread[out_nkhw_thread_desc.GetElementSpace()];
// divide block work
constexpr unsigned NBlockWork =
(out_nkhw_global_desc.GetLength(I0) + NPerBlock - 1) / NPerBlock;
constexpr unsigned KBlockWork =
(out_nkhw_global_desc.GetLength(I1) + KPerBlock - 1) / KPerBlock;
constexpr unsigned HBlockWork =
constexpr index_t NBlockWork = (out_nkhw_global_desc.GetLength(I0) + NPerBlock - 1) / NPerBlock;
constexpr index_t KBlockWork = (out_nkhw_global_desc.GetLength(I1) + KPerBlock - 1) / KPerBlock;
constexpr index_t HBlockWork =
(out_nkhw_global_desc.GetLength(I2) + HoPerBlock - 1) / HoPerBlock;
constexpr unsigned WBlockWork =
constexpr index_t WBlockWork =
(out_nkhw_global_desc.GetLength(I3) + WoPerBlock - 1) / WoPerBlock;
const unsigned block_id = blockIdx.x;
const index_t block_id = blockIdx.x;
unsigned itmp = block_id;
const unsigned n_block_work_id = itmp / (KBlockWork * HBlockWork * WBlockWork);
index_t itmp = block_id;
const index_t n_block_work_id = itmp / (KBlockWork * HBlockWork * WBlockWork);
itmp -= n_block_work_id * (KBlockWork * HBlockWork * WBlockWork);
const unsigned k_block_work_id = itmp / (HBlockWork * WBlockWork);
const index_t k_block_work_id = itmp / (HBlockWork * WBlockWork);
itmp -= k_block_work_id * (HBlockWork * WBlockWork);
const unsigned h_block_work_id = itmp / WBlockWork;
const unsigned w_block_work_id = itmp - h_block_work_id * WBlockWork;
const index_t h_block_work_id = itmp / WBlockWork;
const index_t w_block_work_id = itmp - h_block_work_id * WBlockWork;
const unsigned n_block_data_begin = n_block_work_id * NPerBlock;
const unsigned k_block_data_begin = k_block_work_id * KPerBlock;
const unsigned ho_block_data_begin = h_block_work_id * HoPerBlock;
const unsigned wo_block_data_begin = w_block_work_id * WoPerBlock;
const index_t n_block_data_begin = n_block_work_id * NPerBlock;
const index_t k_block_data_begin = k_block_work_id * KPerBlock;
const index_t ho_block_data_begin = h_block_work_id * HoPerBlock;
const index_t wo_block_data_begin = w_block_work_id * WoPerBlock;
const unsigned hi_block_data_begin = ho_block_data_begin; // minus padding
const unsigned wi_block_data_begin = wo_block_data_begin; // minus padding
const index_t hi_block_data_begin = ho_block_data_begin; // minus padding
const index_t wi_block_data_begin = wo_block_data_begin; // minus padding
// divide thread work
constexpr unsigned NThreadWork = (NPerBlock + NPerThread - 1) / NPerThread;
constexpr unsigned KThreadWork = (KPerBlock + KPerThread - 1) / KPerThread;
constexpr unsigned HThreadWork = (HoPerBlock + HoPerThread - 1) / HoPerThread;
constexpr unsigned WThreadWork = (WoPerBlock + WoPerThread - 1) / WoPerThread;
constexpr index_t NThreadWork = (NPerBlock + NPerThread - 1) / NPerThread;
constexpr index_t KThreadWork = (KPerBlock + KPerThread - 1) / KPerThread;
constexpr index_t HThreadWork = (HoPerBlock + HoPerThread - 1) / HoPerThread;
constexpr index_t WThreadWork = (WoPerBlock + WoPerThread - 1) / WoPerThread;
const unsigned thread_id = threadIdx.x;
const index_t thread_id = threadIdx.x;
itmp = thread_id;
const unsigned n_thread_work_id = itmp / (KThreadWork * HThreadWork * WThreadWork);
const index_t n_thread_work_id = itmp / (KThreadWork * HThreadWork * WThreadWork);
itmp -= n_thread_work_id * (KThreadWork * HThreadWork * WThreadWork);
const unsigned k_thread_work_id = itmp / (HThreadWork * WThreadWork);
const index_t k_thread_work_id = itmp / (HThreadWork * WThreadWork);
itmp -= k_thread_work_id * (HThreadWork * WThreadWork);
const unsigned h_thread_work_id = itmp / WThreadWork;
const unsigned w_thread_work_id = itmp - h_thread_work_id * WThreadWork;
const index_t h_thread_work_id = itmp / WThreadWork;
const index_t w_thread_work_id = itmp - h_thread_work_id * WThreadWork;
const unsigned n_thread_data_begin = n_thread_work_id * NPerThread;
const unsigned k_thread_data_begin = k_thread_work_id * KPerThread;
const unsigned ho_thread_data_begin = h_thread_work_id * HoPerThread;
const unsigned wo_thread_data_begin = w_thread_work_id * WoPerThread;
const index_t n_thread_data_begin = n_thread_work_id * NPerThread;
const index_t k_thread_data_begin = k_thread_work_id * KPerThread;
const index_t ho_thread_data_begin = h_thread_work_id * HoPerThread;
const index_t wo_thread_data_begin = w_thread_work_id * WoPerThread;
const unsigned hi_thread_data_begin = ho_thread_data_begin;
const unsigned wi_thread_data_begin = wo_thread_data_begin;
const index_t hi_thread_data_begin = ho_thread_data_begin;
const index_t wi_thread_data_begin = wo_thread_data_begin;
constexpr auto blockwise_in_copy =
Blockwise4dTensorCopy1<BlockSize,
......@@ -172,7 +170,7 @@ gridwise_direct_convolution_2_nchw_kcyx_nkhw(const Float* const __restrict__ p_i
// set threadwise output tensor to 0
threadwise_4d_tensor_set_zero(out_nkhw_thread_desc, p_out_thread);
for(unsigned c_block_data_begin = 0; c_block_data_begin < C;
for(index_t c_block_data_begin = 0; c_block_data_begin < C;
c_block_data_begin += CPerBlock, __syncthreads())
{
// copy input tensor to LDS
......@@ -191,7 +189,7 @@ gridwise_direct_convolution_2_nchw_kcyx_nkhw(const Float* const __restrict__ p_i
__syncthreads();
for(unsigned c_thread_data = 0; c_thread_data < CPerBlock; c_thread_data += CPerThread)
for(index_t c_thread_data = 0; c_thread_data < CPerBlock; c_thread_data += CPerThread)
{
// threadwise convolution
#if 1
......
src/include/threadwise_2d_tensor_op.hip.hpp
View file @ 766b0a9e
......@@ -16,11 +16,11 @@ __device__ void threadwise_2d_tensor_pointwise_operation_unary(Desc, Float* __re
}
#endif
for(unsigned did0 = 0; did0 < desc.GetLength(I0); ++did0)
for(index_t did0 = 0; did0 < desc.GetLength(I0); ++did0)
{
for(unsigned did1 = 0; did1 < desc.GetLength(I1); ++did1)
for(index_t did1 = 0; did1 < desc.GetLength(I1); ++did1)
{
const unsigned dindex = desc.Get1dIndex(did0, did1);
const index_t dindex = desc.Get1dIndex(did0, did1);
f(p[dindex]);
}
......@@ -47,22 +47,22 @@ __device__ void threadwise_2d_tensor_pointwise_operation_binary_reorder_by_get_d
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr unsigned IR0 = DstFromSrcReorder{}.Get(I0);
constexpr unsigned IR1 = DstFromSrcReorder{}.Get(I1);
constexpr index_t IR0 = DstFromSrcReorder{}.Get(I0);
constexpr index_t IR1 = DstFromSrcReorder{}.Get(I1);
constexpr auto src_desc = SrcDesc{};
constexpr auto dst_desc = DstDesc{};
constexpr auto ref_desc = make_ConstantTensorDescriptor(SrcOpLengths{});
for(unsigned did0 = 0; did0 < ref_desc.GetLength(I0); ++did0)
for(index_t did0 = 0; did0 < ref_desc.GetLength(I0); ++did0)
{
for(unsigned did1 = 0; did1 < ref_desc.GetLength(I1); ++did1)
for(index_t did1 = 0; did1 < ref_desc.GetLength(I1); ++did1)
{
const unsigned aindex = src_desc.Get1dIndex(did0, did1);
const index_t aindex = src_desc.Get1dIndex(did0, did1);
const unsigned did[2] = {did0, did1};
const index_t did[2] = {did0, did1};
const unsigned bindex = dst_desc.Get1dIndex(did[IR0], did[IR1]);
const index_t bindex = dst_desc.Get1dIndex(did[IR0], did[IR1]);
f(p_src[aindex], p_dst[bindex]);
}
......@@ -118,21 +118,21 @@ __device__ void threadwise_2d_tensor_shift_down(Desc, Float* __restrict__ p, IDi
}
#endif
constexpr unsigned nshift = NShift::mValue;
constexpr index_t nshift = NShift::mValue;
constexpr unsigned did0_end =
constexpr index_t did0_end =
is_same<decltype(I0), IDim>::value ? desc.GetLength(I0) - nshift : desc.GetLength(I0);
constexpr unsigned did1_end =
constexpr index_t did1_end =
is_same<decltype(I1), IDim>::value ? desc.GetLength(I1) - nshift : desc.GetLength(I1);
for(unsigned did0 = 0; did0 < did0_end; ++did0)
for(index_t did0 = 0; did0 < did0_end; ++did0)
{
for(unsigned did1 = 0; did1 < did1_end; ++did1)
for(index_t did1 = 0; did1 < did1_end; ++did1)
{
const unsigned dindex = desc.Get1dIndex(did0, did1);
const index_t dindex = desc.Get1dIndex(did0, did1);
const unsigned sindex = dindex + nshift * desc.GetStride(IDim{});
const index_t sindex = dindex + nshift * desc.GetStride(IDim{});
p[dindex] = p[sindex];
}
......
src/include/threadwise_4d_tensor_op.hip.hpp
View file @ 766b0a9e
This diff is collapsed.
src/include/threadwise_direct_convolution.hip.hpp
View file @ 766b0a9e
......@@ -28,28 +28,28 @@ __device__ void threadwise_direct_convolution_1(InDesc,
}
#endif
for(unsigned n = 0; n < out_desc.GetLength(I0); ++n)
for(index_t n = 0; n < out_desc.GetLength(I0); ++n)
{
for(unsigned k = 0; k < out_desc.GetLength(I1); ++k)
for(index_t k = 0; k < out_desc.GetLength(I1); ++k)
{
for(unsigned ho = 0; ho < out_desc.GetLength(I2); ++ho)
for(index_t ho = 0; ho < out_desc.GetLength(I2); ++ho)
{
for(unsigned wo = 0; wo < out_desc.GetLength(I3); ++wo)
for(index_t wo = 0; wo < out_desc.GetLength(I3); ++wo)
{
for(unsigned c = 0; c < wei_desc.GetLength(I1); ++c)
for(index_t c = 0; c < wei_desc.GetLength(I1); ++c)
{
for(unsigned y = 0; y < wei_desc.GetLength(I2); ++y)
for(index_t y = 0; y < wei_desc.GetLength(I2); ++y)
{
for(unsigned x = 0; x < wei_desc.GetLength(I3); ++x)
for(index_t x = 0; x < wei_desc.GetLength(I3); ++x)
{
const unsigned hi = ho + y;
const unsigned wi = wo + x;
const index_t hi = ho + y;
const index_t wi = wo + x;
const unsigned in_index = in_desc.Get1dIndex(n, c, hi, wi);
const index_t in_index = in_desc.Get1dIndex(n, c, hi, wi);
const unsigned wei_index = wei_desc.Get1dIndex(k, c, y, x);
const index_t wei_index = wei_desc.Get1dIndex(k, c, y, x);
const unsigned out_index = out_desc.Get1dIndex(n, k, ho, wo);
const index_t out_index = out_desc.Get1dIndex(n, k, ho, wo);
fused_multiply_accumulate(
p_out[out_index], p_wei[wei_index], p_in[in_index]);
......@@ -125,7 +125,7 @@ __device__ void threadwise_direct_convolution_3(InDesc,
Data p_in_reg[in_reg_desc.GetElementSpace()];
Data p_wei_reg[wei_reg_desc.GetElementSpace()];
constexpr unsigned in_w_new_read = 1;
constexpr index_t in_w_new_read = 1;
constexpr auto in_desc_reg_new_read =
make_ConstantTensorDescriptor(Sequence<in_reg_desc.GetLength(I0),
......@@ -136,7 +136,7 @@ __device__ void threadwise_direct_convolution_3(InDesc,
#if 0
// this verison reused old input data in register, and read new data from LDS
// loop over vertical direction
for(unsigned y = 0; y < wei_desc.GetLength(I2); ++y)
for(index_t y = 0; y < wei_desc.GetLength(I2); ++y)
{
// read first input
threadwise_4d_tensor_copy(in_desc,
......@@ -157,7 +157,7 @@ __device__ void threadwise_direct_convolution_3(InDesc,
in_reg_desc, p_in_reg, wei_reg_desc, p_wei_reg, out_desc, p_out);
// loop over horizontal direction
for(unsigned x = 1; x < wei_desc.GetLength(I3); ++x)
for(index_t x = 1; x < wei_desc.GetLength(I3); ++x)
{
// read new weight
threadwise_4d_tensor_copy(wei_desc,
......@@ -186,10 +186,10 @@ __device__ void threadwise_direct_convolution_3(InDesc,
#elif 1
// this version read all input from LDS when filter moves
// loop over vertical direction
for(unsigned y = 0; y < wei_desc.GetLength(I2); ++y)
for(index_t y = 0; y < wei_desc.GetLength(I2); ++y)
{
// loop over horizontal direction
for(unsigned x = 0; x < wei_desc.GetLength(I3); ++x)
for(index_t x = 0; x < wei_desc.GetLength(I3); ++x)
{
// read new weight
threadwise_4d_tensor_copy(wei_desc,
......
src/include/threadwise_gemm.hip.hpp
View file @ 766b0a9e
This diff is collapsed.
src/include/threadwise_nd_tensor_op.hip.hpp
View file @ 766b0a9e
This diff is collapsed.
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