Skip to content

P
Paddle

ujs_wantz / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle

体验新版 GitCode，发现更多精彩内容 >>

未验证 提交 809a10b6 编写于 作者: F Feiyu Chan 提交者： GitHub
浏览文件
操作

move math_cuda_utils.h to pten/kernels/funcs (#39246)

上级 3e6950d5
隐藏空白更改
内联 并排
Showing with 124 addition and 95 deletion
paddle/fluid/operators/activation_op.cu
浏览文件 @ 809a10b6
......@@ -12,7 +12,6 @@ limitations under the License. */
#include "paddle/fluid/operators/activation_op.h"
#include "paddle/fluid/operators/amp/fp16_type_traits.h"
#include "paddle/fluid/operators/elementwise/elementwise_op_impl.cu.h"
#include "paddle/fluid/operators/math/math_cuda_utils.h"
#include "paddle/fluid/platform/bfloat16.h"
#include "paddle/fluid/platform/device/gpu/gpu_device_function.h"
......
paddle/fluid/operators/interpolate_v2_op.cu
浏览文件 @ 809a10b6
......@@ -12,11 +12,11 @@
#include <algorithm>
#include <string>
#include "paddle/fluid/operators/interpolate_v2_op.h"
#include "paddle/fluid/operators/math/math_cuda_utils.h"
#include "paddle/fluid/platform/device/gpu/gpu_device_function.h"
#include "paddle/fluid/platform/device/gpu/gpu_launch_config.h"
#include "paddle/fluid/platform/device/gpu/gpu_primitives.h"
#include "paddle/fluid/platform/fast_divmod.h"
#include "paddle/pten/kernels/funcs/math_cuda_utils.h"
namespace paddle {
namespace operators {
......@@ -522,7 +522,7 @@ __inline__ __device__ T PartialBlockMin(T val, size_t threads_num_in_block,
if (threadIdx.x < threshold) {
shared_last_idx = (threshold >> 5) - 1;
val = math::warpReduceMin(val, mask);
val = pten::funcs::warpReduceMin(val, mask);
if (lane == 0) {
shared[wid] = val;
}
......@@ -537,7 +537,7 @@ __inline__ __device__ T PartialBlockMin(T val, size_t threads_num_in_block,
if (threadIdx.x < threshold) {
val = (lane <= shared_last_idx) ? shared[lane]
: std::numeric_limits<T>::max();
val = math::warpReduceMin(val, mask);
val = pten::funcs::warpReduceMin(val, mask);
shared_last_val = val;
}
__syncthreads();
......@@ -589,12 +589,15 @@ __global__ void KeBilinearInterpBwShareMemory(
s_data[0][threadIdx.x] = 0.f;
s_data[1][threadIdx.x] = 0.f;
int remain = nthreads - (tid & (-blockDim.x));
int in_top_max_index = math::blockReduceMax(top_right_index, FINAL_MASK);
int in_bot_max_index = math::blockReduceMax(bot_right_index, FINAL_MASK);
int in_top_max_index =
pten::funcs::blockReduceMax(top_right_index, FINAL_MASK);
int in_bot_max_index =
pten::funcs::blockReduceMax(bot_right_index, FINAL_MASK);
if (remain > blockDim.x) {
in_top_min_index = math::blockReduceMin(input_index, FINAL_MASK);
in_bot_min_index = math::blockReduceMin(bot_left_index, FINAL_MASK);
in_top_min_index = pten::funcs::blockReduceMin(input_index, FINAL_MASK);
in_bot_min_index =
pten::funcs::blockReduceMin(bot_left_index, FINAL_MASK);
} else {
in_top_min_index = PartialBlockMin(input_index, remain, FINAL_MASK);
in_bot_min_index = PartialBlockMin(bot_left_index, remain, FINAL_MASK);
......
paddle/fluid/operators/math/bert_encoder_functor.cu
浏览文件 @ 809a10b6
......@@ -18,13 +18,17 @@ limitations under the License. */
#include "paddle/fluid/framework/tensor_util.h"
#include "paddle/fluid/operators/math/bert_encoder_functor.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/math_cuda_utils.h"
#include "paddle/fluid/platform/enforce.h"
#include "paddle/pten/kernels/funcs/math_cuda_utils.h"
namespace paddle {
namespace operators {
namespace math {
// NOTE(chenfeiyu): explicitly use operator+ for float2
// since float2 is not in namespace pten::funcs, ADL won't help
using pten::funcs::operator+;
template <typename T>
__device__ __forceinline__ T local_rsqrt(T num) {
return rsqrt(static_cast<float>(num));
......@@ -34,11 +38,12 @@ __device__ __forceinline__ half local_rsqrt(half num) { return hrsqrt(num); }
#endif
template <typename T, int TPB>
__device__ inline void LayerNormSmall(T val, const kvp<T> &thread_data,
__device__ inline void LayerNormSmall(T val,
const pten::funcs::kvp<T> &thread_data,
const int ld, const int idx,
const float *bias, const float *scale,
T *output, T eps) {
using BlockReduce = cub::BlockReduce<kvp<T>, TPB>;
using BlockReduce = cub::BlockReduce<pten::funcs::kvp<T>, TPB>;
__shared__ typename BlockReduce::TempStorage temp_storage;
__shared__ T mu; // mean
__shared__ T rsigma; // 1 / std.dev.
......@@ -59,10 +64,11 @@ __device__ inline void LayerNormSmall(T val, const kvp<T> &thread_data,
}
template <typename T, int TPB>
__device__ inline void LayerNorm(const kvp<T> &thread_data, const int ld,
const int offset, const float *bias,
const float *scale, T *output, T eps) {
using BlockReduce = cub::BlockReduce<kvp<T>, TPB>;
__device__ inline void LayerNorm(const pten::funcs::kvp<T> &thread_data,
const int ld, const int offset,
const float *bias, const float *scale,
T *output, T eps) {
using BlockReduce = cub::BlockReduce<pten::funcs::kvp<T>, TPB>;
__shared__ typename BlockReduce::TempStorage temp_storage;
__shared__ T mu; // mean
__shared__ T rsigma; // 1 / std.dev.
......@@ -85,10 +91,11 @@ __device__ inline void LayerNorm(const kvp<T> &thread_data, const int ld,
}
template <typename T, typename T2, int TPB>
__device__ inline void LayerNorm2(const kvp<T> &thread_data, const int ld,
const int offset, const float2 *bias,
const float2 *scale, T2 *output, T eps) {
using BlockReduce = cub::BlockReduce<kvp<T>, TPB>;
__device__ inline void LayerNorm2(const pten::funcs::kvp<T> &thread_data,
const int ld, const int offset,
const float2 *bias, const float2 *scale,
T2 *output, T eps) {
using BlockReduce = cub::BlockReduce<pten::funcs::kvp<T>, TPB>;
__shared__ typename BlockReduce::TempStorage temp_storage;
__shared__ T mu; // mean
__shared__ T rsigma; // 1 / std.dev.
......@@ -137,7 +144,7 @@ __global__ void EmbEltwiseLayernormKernel(int hidden, const int64_t *ids,
const int64_t out_offset = seq_pos * hidden;
kvp<T> thread_data(0, 0);
pten::funcs::kvp<T> thread_data(0, 0);
#pragma unroll
for (int it = threadIdx.x; it < hidden; it += TPB) {
......@@ -148,7 +155,8 @@ __global__ void EmbEltwiseLayernormKernel(int hidden, const int64_t *ids,
output[out_offset + it] = val;
const T rhiddenval = rhidden * val;
thread_data = pair_sum(thread_data, kvp<T>(rhiddenval, rhiddenval * val));
thread_data = pair_sum(thread_data,
pten::funcs::kvp<T>(rhiddenval, rhiddenval * val));
}
LayerNorm<T, TPB>(thread_data, hidden, out_offset, bias, scale, output, eps);
}
......@@ -180,7 +188,7 @@ __global__ void EmbEltwiseLayernormKernel<half, 256>(
const int64_t out_offset = seq_pos * hidden;
kvp<half> thread_data(0, 0);
pten::funcs::kvp<half> thread_data(0, 0);
#pragma unroll
for (int it = threadIdx.x; it < hidden; it += 256) {
......@@ -191,8 +199,8 @@ __global__ void EmbEltwiseLayernormKernel<half, 256>(
output[out_offset + it] = val;
const half rhiddenval = rhidden * val;
thread_data =
pair_sum(thread_data, kvp<half>(rhiddenval, rhiddenval * val));
thread_data = pair_sum(
thread_data, pten::funcs::kvp<half>(rhiddenval, rhiddenval * val));
}
LayerNorm<half, 256>(thread_data, hidden, out_offset, bias, scale, output,
eps);
......@@ -233,10 +241,10 @@ __global__ void SoftmaxKernelWithEltadd(T *qk_buf_, const T *bias_qk_,
? static_cast<float>(qk_buf_[threadIdx.x + qk_offset] +
bias_qk_[threadIdx.x + qk_offset])
: -1e20f;
float max_val = blockReduceMax<float>(tmp, mask);
float max_val = pten::funcs::blockReduceMax<float>(tmp, mask);
float qk_tmp = threadIdx.x < seq_len ? __expf(tmp - max_val) : 0.0f;
float sum_val = blockReduceSum<float>(qk_tmp, mask);
float sum_val = pten::funcs::blockReduceSum<float>(qk_tmp, mask);
if (threadIdx.x < seq_len)
qk_buf_[threadIdx.x + qk_offset] = (T)(qk_tmp / sum_val);
......@@ -256,10 +264,10 @@ __global__ void SoftmaxKernelWithEltadd<half>(
? static_cast<float>(qk_buf_[threadIdx.x + qk_offset] +
bias_qk_[threadIdx.x + qk_offset])
: -1e20f;
float max_val = blockReduceMax<float>(tmp, mask);
float max_val = pten::funcs::blockReduceMax<float>(tmp, mask);
float qk_tmp = threadIdx.x < seq_len ? __expf(tmp - max_val) : 0.0f;
float sum_val = blockReduceSum<float>(qk_tmp, mask);
float sum_val = pten::funcs::blockReduceSum<float>(qk_tmp, mask);
if (threadIdx.x < seq_len)
qk_buf_[threadIdx.x + qk_offset] = (half)(qk_tmp / sum_val);
......@@ -276,19 +284,20 @@ __global__ void SoftmaxKernelWithEltadd2(T *qk_buf_, const T *bias_qk_,
int idx = threadIdx.x;
assert(blockDim.x % 32 == 0);
float2 tmp =
idx < seq_len
? ToFloat2<T>(qk_buf_[idx + qk_offset] + bias_qk_[idx + qk_offset])
: make_float2(-1e20f, -1e20f);
float max_val = blockReduceMax<float>(max(tmp.x, tmp.y), mask);
float2 tmp = idx < seq_len
? pten::funcs::ToFloat2<T>(qk_buf_[idx + qk_offset] +
bias_qk_[idx + qk_offset])
: make_float2(-1e20f, -1e20f);
float max_val = pten::funcs::blockReduceMax<float>(max(tmp.x, tmp.y), mask);
float2 qk_tmp = idx < seq_len ? make_float2(__expf(tmp.x - max_val),
__expf(tmp.y - max_val))
: make_float2(0.f, 0.f);
float sum_val = blockReduceSum<float>(qk_tmp.x + qk_tmp.y, mask) + 1e-6f;
float sum_val =
pten::funcs::blockReduceSum<float>(qk_tmp.x + qk_tmp.y, mask) + 1e-6f;
if (idx < seq_len) {
qk_buf_[idx + qk_offset] =
FloatsToPair<T>(qk_tmp.x / sum_val, qk_tmp.y / sum_val);
pten::funcs::FloatsToPair<T>(qk_tmp.x / sum_val, qk_tmp.y / sum_val);
}
}
......@@ -304,18 +313,20 @@ __global__ void SoftmaxKernelWithEltadd2<half2>(
int idx = threadIdx.x;
assert(blockDim.x % 32 == 0);
float2 tmp = idx < seq_len ? ToFloat2<half2>(qk_buf_[idx + qk_offset] +
bias_qk_[idx + qk_offset])
: make_float2(-1e20f, -1e20f);
float max_val = blockReduceMax<float>(max(tmp.x, tmp.y), mask);
float2 tmp = idx < seq_len
? pten::funcs::ToFloat2<half2>(qk_buf_[idx + qk_offset] +
bias_qk_[idx + qk_offset])
: make_float2(-1e20f, -1e20f);
float max_val = pten::funcs::blockReduceMax<float>(max(tmp.x, tmp.y), mask);
float2 qk_tmp = idx < seq_len ? make_float2(__expf(tmp.x - max_val),
__expf(tmp.y - max_val))
: make_float2(0.f, 0.f);
float sum_val = blockReduceSum<float>(qk_tmp.x + qk_tmp.y, mask) + 1e-6f;
float sum_val =
pten::funcs::blockReduceSum<float>(qk_tmp.x + qk_tmp.y, mask) + 1e-6f;
if (idx < seq_len) {
qk_buf_[idx + qk_offset] =
FloatsToPair<half2>(qk_tmp.x / sum_val, qk_tmp.y / sum_val);
qk_buf_[idx + qk_offset] = pten::funcs::FloatsToPair<half2>(
qk_tmp.x / sum_val, qk_tmp.y / sum_val);
}
#endif
}
......@@ -338,14 +349,14 @@ __global__ void SoftmaxKernelWithEltaddForLarge(T *qk_buf, const T *bias_qk,
bias_qk[threadIdx.x + i + qk_offset]
: stride_max;
}
T max_val = blockReduceMax<T>(stride_max, mask);
T max_val = pten::funcs::blockReduceMax<T>(stride_max, mask);
T stride_sum = 0.f;
for (int i = 0; i < seq_len; i += blockDim.x) {
stride_sum += __expf(qk_buf[threadIdx.x + i + qk_offset] +
bias_qk[threadIdx.x + i + qk_offset] - max_val);
}
T sum_val = blockReduceSum<T>(stride_sum, mask);
T sum_val = pten::funcs::blockReduceSum<T>(stride_sum, mask);
for (int i = 0; i < seq_len; i += blockDim.x) {
qk_buf[threadIdx.x + i + qk_offset] =
......@@ -371,7 +382,7 @@ __global__ void SoftmaxKernelWithEltaddForLarge(
bias_qk[threadIdx.x + i + qk_offset]);
stride_max = tmp > stride_max ? tmp : stride_max;
}
float max_val = blockReduceMax<float>(stride_max, mask);
float max_val = pten::funcs::blockReduceMax<float>(stride_max, mask);
float stride_sum = 0.f;
for (int i = 0; i < seq_len; i += blockDim.x) {
......@@ -379,7 +390,7 @@ __global__ void SoftmaxKernelWithEltaddForLarge(
bias_qk[threadIdx.x + i + qk_offset]);
stride_sum += __expf(tmp - max_val);
}
float sum_val = blockReduceSum<float>(stride_sum, mask);
float sum_val = pten::funcs::blockReduceSum<float>(stride_sum, mask);
for (int i = 0; i < seq_len; i += blockDim.x) {
float tmp =
......@@ -403,28 +414,33 @@ __global__ void SoftmaxKernelWithEltaddForLarge2(T *qk_buf_, const T *bias_qk_,
float2 stride_max = make_float2(-1e20f, -1e20f);
for (int i = 0; i < seq_len; i += blockDim.x) {
float2 cur = ToFloat2<T>(qk_buf_[threadIdx.x + i + qk_offset] +
bias_qk_[threadIdx.x + i + qk_offset]);
float2 cur =
pten::funcs::ToFloat2<T>(qk_buf_[threadIdx.x + i + qk_offset] +
bias_qk_[threadIdx.x + i + qk_offset]);
stride_max.x = max(stride_max.x, cur.x);
stride_max.y = max(stride_max.y, cur.y);
}
float max_val = blockReduceMax<float>(max(stride_max.x, stride_max.y), mask);
float max_val =
pten::funcs::blockReduceMax<float>(max(stride_max.x, stride_max.y), mask);
float2 stride_sum = make_float2(0.f, 0.f);
for (int i = 0; i < seq_len; i += blockDim.x) {
float2 cur = ToFloat2<T>(qk_buf_[threadIdx.x + i + qk_offset] +
bias_qk_[threadIdx.x + i + qk_offset]);
float2 cur =
pten::funcs::ToFloat2<T>(qk_buf_[threadIdx.x + i + qk_offset] +
bias_qk_[threadIdx.x + i + qk_offset]);
stride_sum.x += __expf(cur.x - max_val);
stride_sum.y += __expf(cur.y - max_val);
}
float sum_val =
blockReduceSum<float>(stride_sum.x + stride_sum.y, mask) + 1e-6f;
pten::funcs::blockReduceSum<float>(stride_sum.x + stride_sum.y, mask) +
1e-6f;
for (int i = 0; i < seq_len; i += blockDim.x) {
float2 cur = ToFloat2<T>(qk_buf_[threadIdx.x + i + qk_offset] +
bias_qk_[threadIdx.x + i + qk_offset]);
qk_buf_[threadIdx.x + i + qk_offset] = FloatsToPair<T>(
float2 cur =
pten::funcs::ToFloat2<T>(qk_buf_[threadIdx.x + i + qk_offset] +
bias_qk_[threadIdx.x + i + qk_offset]);
qk_buf_[threadIdx.x + i + qk_offset] = pten::funcs::FloatsToPair<T>(
__expf(cur.x - max_val) / sum_val, __expf(cur.y - max_val) / sum_val);
}
}
......@@ -443,28 +459,33 @@ __global__ void SoftmaxKernelWithEltaddForLarge2(
float2 stride_max = make_float2(-1e20f, -1e20f);
for (int i = 0; i < seq_len; i += blockDim.x) {
float2 cur = ToFloat2<half2>(qk_buf_[threadIdx.x + i + qk_offset] +
bias_qk_[threadIdx.x + i + qk_offset]);
float2 cur =
pten::funcs::ToFloat2<half2>(qk_buf_[threadIdx.x + i + qk_offset] +
bias_qk_[threadIdx.x + i + qk_offset]);
stride_max.x = max(stride_max.x, cur.x);
stride_max.y = max(stride_max.y, cur.y);
}
float max_val = blockReduceMax<float>(max(stride_max.x, stride_max.y), mask);
float max_val =
pten::funcs::blockReduceMax<float>(max(stride_max.x, stride_max.y), mask);
float2 stride_sum = make_float2(0.f, 0.f);
for (int i = 0; i < seq_len; i += blockDim.x) {
float2 cur = ToFloat2<half2>(qk_buf_[threadIdx.x + i + qk_offset] +
bias_qk_[threadIdx.x + i + qk_offset]);
float2 cur =
pten::funcs::ToFloat2<half2>(qk_buf_[threadIdx.x + i + qk_offset] +
bias_qk_[threadIdx.x + i + qk_offset]);
stride_sum.x += __expf(cur.x - max_val);
stride_sum.y += __expf(cur.y - max_val);
}
float sum_val =
blockReduceSum<float>(stride_sum.x + stride_sum.y, mask) + 1e-6f;
pten::funcs::blockReduceSum<float>(stride_sum.x + stride_sum.y, mask) +
1e-6f;
for (int i = 0; i < seq_len; i += blockDim.x) {
float2 cur = ToFloat2<half2>(qk_buf_[threadIdx.x + i + qk_offset] +
bias_qk_[threadIdx.x + i + qk_offset]);
qk_buf_[threadIdx.x + i + qk_offset] = FloatsToPair<half2>(
float2 cur =
pten::funcs::ToFloat2<half2>(qk_buf_[threadIdx.x + i + qk_offset] +
bias_qk_[threadIdx.x + i + qk_offset]);
qk_buf_[threadIdx.x + i + qk_offset] = pten::funcs::FloatsToPair<half2>(
__expf(cur.x - max_val) / sum_val, __expf(cur.y - max_val) / sum_val);
}
#endif
......@@ -595,13 +616,14 @@ __global__ void SkipLayerNormSmallKernel(int num, int hidden, const T *input1,
const T rld = T(1) / T(hidden);
const int offset = blockIdx.x * hidden;
cub::Sum pair_sum;
kvp<T> thread_data(0, 0);
pten::funcs::kvp<T> thread_data(0, 0);
const int idx = offset + threadIdx.x;
T val = 0;
if (threadIdx.x < hidden) {
val = input1[idx] + input2[idx];
const T rldval = rld * val;
thread_data = pair_sum(thread_data, kvp<T>(rldval, rldval * val));
thread_data =
pair_sum(thread_data, pten::funcs::kvp<T>(rldval, rldval * val));
}
LayerNormSmall<T, TPB>(val, thread_data, hidden, idx, bias, scale, output,
eps);
......@@ -617,13 +639,14 @@ __global__ void SkipLayerNormSmallKernel<half, 32>(
const half rld = half(1) / half(hidden);
const int offset = blockIdx.x * hidden;
cub::Sum pair_sum;
kvp<half> thread_data(0, 0);
pten::funcs::kvp<half> thread_data(0, 0);
const int idx = offset + threadIdx.x;
half val = 0;
if (threadIdx.x < hidden) {
val = input1[idx] + input2[idx];
const half rldval = rld * val;
thread_data = pair_sum(thread_data, kvp<half>(rldval, rldval * val));
thread_data =
pair_sum(thread_data, pten::funcs::kvp<half>(rldval, rldval * val));
}
LayerNormSmall<half, 32>(val, thread_data, hidden, idx, bias, scale, output,
eps);
......@@ -638,13 +661,14 @@ __global__ void SkipLayerNormSmallKernel<half, 128>(
const half rld = half(1) / half(hidden);
const int offset = blockIdx.x * hidden;
cub::Sum pair_sum;
kvp<half> thread_data(0, 0);
pten::funcs::kvp<half> thread_data(0, 0);
const int idx = offset + threadIdx.x;
half val = 0;
if (threadIdx.x < hidden) {
val = input1[idx] + input2[idx];
const half rldval = rld * val;
thread_data = pair_sum(thread_data, kvp<half>(rldval, rldval * val));
thread_data =
pair_sum(thread_data, pten::funcs::kvp<half>(rldval, rldval * val));
}
LayerNormSmall<half, 128>(val, thread_data, hidden, idx, bias, scale, output,
eps);
......@@ -659,13 +683,14 @@ __global__ void SkipLayerNormSmallKernel<half, 384>(
const half rld = half(1) / half(hidden);
const int offset = blockIdx.x * hidden;
cub::Sum pair_sum;
kvp<half> thread_data(0, 0);
pten::funcs::kvp<half> thread_data(0, 0);
const int idx = offset + threadIdx.x;
half val = 0;
if (threadIdx.x < hidden) {
val = input1[idx] + input2[idx];
const half rldval = rld * val;
thread_data = pair_sum(thread_data, kvp<half>(rldval, rldval * val));
thread_data =
pair_sum(thread_data, pten::funcs::kvp<half>(rldval, rldval * val));
}
LayerNormSmall<half, 384>(val, thread_data, hidden, idx, bias, scale, output,
eps);
......@@ -681,13 +706,14 @@ __global__ void SkipLayerNormKernel(int num, int hidden, const T *input1,
const T rld = T(1) / T(hidden);
const int offset = blockIdx.x * hidden;
cub::Sum pair_sum;
kvp<T> thread_data(0, 0);
pten::funcs::kvp<T> thread_data(0, 0);
for (int it = threadIdx.x; it < hidden; it += TPB) {
const int idx = offset + it;
const T val = input1[idx] + input2[idx];
const T rldval = rld * val;
thread_data = pair_sum(thread_data, kvp<T>(rldval, rldval * val));
thread_data =
pair_sum(thread_data, pten::funcs::kvp<T>(rldval, rldval * val));
output[idx] = val;
}
LayerNorm<T, TPB>(thread_data, hidden, offset, bias, scale, output, eps);
......@@ -705,13 +731,14 @@ __global__ void SkipLayerNormKernel<half, 256>(int num, int hidden,
const half rld = half(1) / half(hidden);
const int offset = blockIdx.x * hidden;
cub::Sum pair_sum;
kvp<half> thread_data(0, 0);
pten::funcs::kvp<half> thread_data(0, 0);
for (int it = threadIdx.x; it < hidden; it += 256) {
const int idx = offset + it;
const half val = input1[idx] + input2[idx];
const half rldval = rld * val;
thread_data = pair_sum(thread_data, kvp<half>(rldval, rldval * val));
thread_data =
pair_sum(thread_data, pten::funcs::kvp<half>(rldval, rldval * val));
output[idx] = val;
}
LayerNorm<half, 256>(thread_data, hidden, offset, bias, scale, output, eps);
......@@ -727,13 +754,14 @@ __global__ void SkipLayerNormKernel2(int num, int hidden, const T2 *input1,
const T rld = T(0.5f / hidden); // because hidden is hidden/2
const int offset = blockIdx.x * hidden;
cub::Sum pair_sum;
kvp<T> thread_data(0, 0);
pten::funcs::kvp<T> thread_data(0, 0);
for (int it = threadIdx.x; it < hidden; it += TPB) {
const int idx = offset + it;
const T2 val2 = input1[idx] + input2[idx];
thread_data = pair_sum(
thread_data, kvp<T>(rld * (val2.x + val2.y),
thread_data,
pten::funcs::kvp<T>(rld * (val2.x + val2.y),
rld * val2.x * val2.x + rld * val2.y * val2.y));
output[idx] = val2;
}
......@@ -751,13 +779,14 @@ __global__ void SkipLayerNormKernel2<half, half2, 256>(
const half rld = half(0.5f / hidden); // because hidden is hidden/2
const int offset = blockIdx.x * hidden;
cub::Sum pair_sum;
kvp<half> thread_data(0, 0);
pten::funcs::kvp<half> thread_data(0, 0);
for (int it = threadIdx.x; it < hidden; it += 256) {
const int idx = offset + it;
const half2 val2 = input1[idx] + input2[idx];
thread_data = pair_sum(
thread_data, kvp<half>(rld * (val2.x + val2.y),
thread_data,
pten::funcs::kvp<half>(rld * (val2.x + val2.y),
rld * val2.x * val2.x + rld * val2.y * val2.y));
output[idx] = val2;
}
......
paddle/fluid/operators/optimizers/lars_momentum_op.cu
浏览文件 @ 809a10b6
......@@ -14,9 +14,9 @@ limitations under the License. */
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/amp/fp16_type_traits.h"
#include "paddle/fluid/operators/math/math_cuda_utils.h"
#include "paddle/fluid/operators/optimizers/lars_momentum_op.h"
#include "paddle/fluid/platform/fast_divmod.h"
#include "paddle/pten/kernels/funcs/math_cuda_utils.h"
#if CUDA_VERSION >= 11000
#include <cooperative_groups.h>
......@@ -170,8 +170,8 @@ __global__ void L2NormKernel(
g_tmp += (tmp1 * tmp1);
tid += grid_stride;
}
p_tmp = math::blockReduceSum<MT>(p_tmp, FINAL_MASK);
g_tmp = math::blockReduceSum<MT>(g_tmp, FINAL_MASK);
p_tmp = pten::funcs::blockReduceSum<MT>(p_tmp, FINAL_MASK);
g_tmp = pten::funcs::blockReduceSum<MT>(g_tmp, FINAL_MASK);
if (threadIdx.x == 0) {
p_buffer[blockIdx.x] = p_tmp;
......@@ -181,8 +181,8 @@ __global__ void L2NormKernel(
cg->sync(); // Grid sync for writring partial result to gloabl memory
MT p_part_sum = threadIdx.x < gridDim.x ? p_buffer[threadIdx.x] : 0;
MT g_part_sum = threadIdx.x < gridDim.x ? g_buffer[threadIdx.x] : 0;
MT tmp0 = math::blockReduceSum<MT>(p_part_sum, FINAL_MASK);
MT tmp1 = math::blockReduceSum<MT>(g_part_sum, FINAL_MASK);
MT tmp0 = pten::funcs::blockReduceSum<MT>(p_part_sum, FINAL_MASK);
MT tmp1 = pten::funcs::blockReduceSum<MT>(g_part_sum, FINAL_MASK);
if (threadIdx.x == 0) {
s_buffer[0] = tmp0;
s_buffer[1] = tmp1;
......@@ -294,9 +294,10 @@ __global__ void MomentumLarsKernel(
MT param_part_norm = threadIdx.x < thresh ? p_buffer[threadIdx.x] : 0;
MT grad_part_norm = threadIdx.x < thresh ? g_buffer[threadIdx.x] : 0;
__syncthreads();
MT param_norm = Sqrt(math::blockReduceSum<MT>(param_part_norm, FINAL_MASK));
MT grad_norm = Sqrt(rescale_grad_pow *
math::blockReduceSum<MT>(grad_part_norm, FINAL_MASK));
MT param_norm =
Sqrt(pten::funcs::blockReduceSum<MT>(param_part_norm, FINAL_MASK));
MT grad_norm = Sqrt(rescale_grad_pow * pten::funcs::blockReduceSum<MT>(
grad_part_norm, FINAL_MASK));
#endif
MomentumUpdate<T, MT>(param, grad, velocity, param_out, velocity_out,
master_param, master_param_out, learning_rate, mu,
......
paddle/fluid/operators/softmax_cudnn_op.cu.h
浏览文件 @ 809a10b6
......@@ -16,7 +16,6 @@ limitations under the License. */
#include "paddle/fluid/operators/amp/fp16_type_traits.h"
#include "paddle/fluid/operators/kernel_primitives/kernel_primitives.h"
#include "paddle/fluid/operators/math/math_cuda_utils.h"
#include "paddle/fluid/operators/softmax_op.h"
#include "paddle/fluid/platform/device/gpu/gpu_device_function.h"
#include "paddle/fluid/platform/device/gpu/gpu_dnn.h"
......
paddle/fluid/operators/math/math_cuda_utils.h → paddle/pten/kernels/funcs/math_cuda_utils.h
浏览文件 @ 809a10b6
......@@ -23,9 +23,8 @@ limitations under the License. */
#include <algorithm>
namespace paddle {
namespace operators {
namespace math {
namespace pten {
namespace funcs {
template <typename T>
__device__ __forceinline__ T FromFloat(float a);
......@@ -315,6 +314,5 @@ __inline__ __device__ T PartialBlockReduceMin(T val, unsigned mask) {
return val;
}
} // namespace math
} // namespace operators
} // namespace paddle
} // namespace funcs
} // namespace pten
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或