Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Fix quant and dequant cuda kernels when quant_axis==1 #40772

Merged
merged 3 commits into from
Mar 23, 2022
Merged

Fix quant and dequant cuda kernels when quant_axis==1 #40772

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
d44a952
Fix fake quant and dequant when quant_axis=1
wanghaoshuang Mar 18, 2022
9ccf69f
Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…
wanghaoshuang Mar 18, 2022
0d9e0bb
Fix quant and dequant for quan_axis != 0
wanghaoshuang Mar 21, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
50 changes: 29 additions & 21 deletions paddle/fluid/operators/fake_dequantize_op.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -58,19 +58,15 @@ __global__ void DequantizeOneScaleQuantAxis0(const T* in, const T* scale,
}

template <typename T>
__global__ void DequantizeOneScaleQuantAxis1(const T* in, const T* scale,
T max_range, const int num,
const int cin, const int cout,
T* out) {
int bid = blockIdx.x;
T s = scale[bid % cout];

int wh_size = num / (cin * cout);
const T* in_current = in + bid * wh_size;
T* out_current = out + bid * wh_size;

for (int i = threadIdx.x; i < wh_size; i += blockDim.x) {
out_current[i] = in_current[i] * s / max_range;
__global__ void DequantizeOneScaleQuantAxisN(const T* in, const T* scale,
const T max_range,
const int64_t num,
const int n_scales,
const int quant_stride, T* out) {
int64_t idx = blockDim.x * blockIdx.x + threadIdx.x;
for (int64_t i = idx; i < num; i += blockDim.x * gridDim.x) {
T s = scale[(i / quant_stride) % n_scales];
out[i] = in[i] * s / max_range;
}
}

Expand Down Expand Up @@ -98,20 +94,32 @@ struct ChannelDequantizeFunctor<platform::CUDADeviceContext, T> {
const T* in_data = in->data<T>();
T* out_data = out->mutable_data<T>(dev_ctx.GetPlace());
if (scale_num == 1) {
int num = in->numel();
int64_t num = in->numel();
const T* scale_factor = scales[0]->data<T>();
if (quant_axis == 0) {
int grid = in_dims[0];
int block = 1024;
DequantizeOneScaleQuantAxis0<T><<<grid, block, 0, dev_ctx.stream()>>>(
in_data, scale_factor, max_range, num, in_dims[0], out_data);
} else if (quant_axis == 1) {
// Dequantize weight of Cin * Cout * W * H
int grid = in_dims[0] * in_dims[1];
int block = 1024;
DequantizeOneScaleQuantAxis1<T><<<grid, block, 0, dev_ctx.stream()>>>(
in_data, scale_factor, max_range, num, in_dims[0], in_dims[1],
out_data);
} else {
int quant_stride = 1;
for (int i = quant_axis + 1; i < in_dims.size(); i++) {
quant_stride *= in_dims[i];
}

int64_t block_size = std::min(
num, static_cast<int64_t>(dev_ctx.GetMaxThreadsPerBlock() / 4));
int64_t max_threads =
dev_ctx.GetMaxPhysicalThreadCount(); // SM * block_per_SM
const int64_t max_blocks = std::max(
((max_threads - 1) / block_size + 1), static_cast<int64_t>(1));
const int64_t grid_size =
std::min(max_blocks, (num + block_size - 1) / block_size);

DequantizeOneScaleQuantAxisN<
T><<<grid_size, block_size, 0, dev_ctx.stream()>>>(
in_data, scale_factor, max_range, num, in_dims[quant_axis],
quant_stride, out_data);
}
} else if (scale_num == 2) {
// Not need to consider quant_axis
Expand Down
58 changes: 33 additions & 25 deletions paddle/fluid/operators/fake_quantize_op.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -273,18 +273,18 @@ struct ClipAndFakeQuantDequantFunctor<platform::CUDADeviceContext, T> {
template <typename T>
__global__ void ChannelClipAndQuantKernelQuantAxis0(const T* in, const T* scale,
const int bin_cnt,
const int n, const int c,
T* out) {
const int64_t n,
const int c, T* out) {
int tid = threadIdx.x;

int channel_size = n / c;
int64_t channel_size = n / c;
const T* in_c = in + blockIdx.x * channel_size;
T* out_c = out + blockIdx.x * channel_size;

T s = scale[blockIdx.x];
T inv_s = inverse(s);

for (int i = tid; i < channel_size; i += blockDim.x) {
for (int64_t i = tid; i < channel_size; i += blockDim.x) {
T x = in_c[i];
T v = x > s ? s : x;
v = v < -s ? -s : v;
Expand All @@ -293,25 +293,20 @@ __global__ void ChannelClipAndQuantKernelQuantAxis0(const T* in, const T* scale,
}
}

// ChannelClipAndQuantKernel for quant_axis is 1
// ChannelClipAndQuantKernel for quant_axis is N
template <typename T>
__global__ void ChannelClipAndQuantKernelQuantAxis1(const T* in, const T* scale,
const int bin_cnt,
const int n, const int cin,
const int cout, T* out) {
T s = scale[blockIdx.x % cout];
T inv_s = inverse(s);

int wh_size = n / (cin * cout);
const T* in_c = in + blockIdx.x * wh_size;
T* out_c = out + blockIdx.x * wh_size;

for (int i = threadIdx.x; i < wh_size; i += blockDim.x) {
T x = in_c[i];
__global__ void ChannelClipAndQuantKernelQuantAxisN(
const T* in, const T* scale, const int bin_cnt, const int64_t n,
const int nScale, const int quant_stride, T* out) {
int64_t idx = blockDim.x * blockIdx.x + threadIdx.x;
for (int64_t i = idx; i < n; i += blockDim.x * gridDim.x) {
T s = scale[(i / quant_stride) % nScale];
T inv_s = 1.0 / s;
T x = in[i];
T v = x > s ? s : x;
v = v < -s ? -s : v;
v = bin_cnt * inv_s * v;
out_c[i] = round(v);
out[i] = round(v);
}
}

Expand All @@ -327,7 +322,7 @@ struct ChannelClipAndFakeQuantFunctor<platform::CUDADeviceContext, T> {
"the received is %d",
quant_axis));

int num = in.numel();
int64_t num = in.numel();
auto in_dims = in.dims();
const T* in_data = in.data<T>();
const T* scale_data = scale.data<T>();
Expand All @@ -338,11 +333,24 @@ struct ChannelClipAndFakeQuantFunctor<platform::CUDADeviceContext, T> {
int block = 1024;
ChannelClipAndQuantKernelQuantAxis0<T><<<grid, block, 0, ctx.stream()>>>(
in_data, scale_data, bin_cnt, num, in_dims[0], out_data);
} else if (quant_axis == 1) {
int grid = in_dims[0] * in_dims[1];
int block = 1024;
ChannelClipAndQuantKernelQuantAxis1<T><<<grid, block, 0, ctx.stream()>>>(
in_data, scale_data, bin_cnt, num, in_dims[0], in_dims[1], out_data);
} else {
int quant_stride = 1;
for (int i = quant_axis + 1; i < in_dims.size(); i++) {
quant_stride *= in_dims[i];
}
int64_t block_size =
std::min(num, static_cast<int64_t>(ctx.GetMaxThreadsPerBlock() / 4));
int64_t max_threads =
ctx.GetMaxPhysicalThreadCount(); // SM * block_per_SM
const int64_t max_blocks = std::max(((max_threads - 1) / block_size + 1),
static_cast<int64_t>(1));

const int64_t grid_size =
std::min(max_blocks, (num + block_size - 1) / block_size);

ChannelClipAndQuantKernelQuantAxisN<T><<<grid_size, block_size>>>(
in_data, scale_data, bin_cnt, num, in_dims[quant_axis], quant_stride,
out_data);
}
}
};
Expand Down