opencv/modules/dnn/src/vkcom/shader/softmax.comp
WuZhiwen 6e3ea8b49d Merge pull request #12703 from wzw-intel:vkcom
* dnn: Add a Vulkan based backend

This commit adds a new backend "DNN_BACKEND_VKCOM" and a
new target "DNN_TARGET_VULKAN". VKCOM means vulkan based
computation library.

This backend uses Vulkan API and SPIR-V shaders to do
the inference computation for layers. The layer types
that implemented in DNN_BACKEND_VKCOM include:
Conv, Concat, ReLU, LRN, PriorBox, Softmax, MaxPooling,
AvePooling, Permute

This is just a beginning work for Vulkan in OpenCV DNN,
more layer types will be supported and performance
tuning is on the way.

Signed-off-by: Wu Zhiwen <zhiwen.wu@intel.com>

* dnn/vulkan: Add FindVulkan.cmake to detect Vulkan SDK

In order to build dnn with Vulkan support, need installing
Vulkan SDK and setting environment variable "VULKAN_SDK" and
add "-DWITH_VULKAN=ON" to cmake command.

You can download Vulkan SDK from:
https://vulkan.lunarg.com/sdk/home#linux

For how to install, see
https://vulkan.lunarg.com/doc/sdk/latest/linux/getting_started.html
https://vulkan.lunarg.com/doc/sdk/latest/windows/getting_started.html
https://vulkan.lunarg.com/doc/sdk/latest/mac/getting_started.html
respectively for linux, windows and mac.

To run the vulkan backend, also need installing mesa driver.
On Ubuntu, use this command 'sudo apt-get install mesa-vulkan-drivers'

To test, use command '$BUILD_DIR/bin/opencv_test_dnn --gtest_filter=*VkCom*'

Signed-off-by: Wu Zhiwen <zhiwen.wu@intel.com>

* dnn/Vulkan: dynamically load Vulkan runtime

No compile-time dependency on Vulkan library.
If Vulkan runtime is unavailable, fallback to CPU path.

Use environment "OPENCL_VULKAN_RUNTIME" to specify path to your
own vulkan runtime library.

Signed-off-by: Wu Zhiwen <zhiwen.wu@intel.com>

* dnn/Vulkan: Add a python script to compile GLSL shaders to SPIR-V shaders

The SPIR-V shaders are in format of text-based 32-bit hexadecimal
numbers, and inserted into .cpp files as unsigned int32 array.

* dnn/Vulkan: Put Vulkan headers into 3rdparty directory and some other fixes

Vulkan header files are copied from
https://github.com/KhronosGroup/Vulkan-Docs/tree/master/include/vulkan
to 3rdparty/include

Fix the Copyright declaration issue.

Refine OpenCVDetectVulkan.cmake

* dnn/Vulkan: Add vulkan backend tests into existing ones.

Also fixed some test failures.

- Don't use bool variable as uniform for shader
- Fix dispathed group number beyond max issue
- Bypass "group > 1" convolution. This should be support in future.

* dnn/Vulkan: Fix multiple initialization in one thread.
2018-10-29 17:51:26 +03:00

79 lines
2.1 KiB
Plaintext

#version 450
#define LOCAL_SZ_X 256
layout(binding = 0) readonly buffer buf0{
float input_buffer[]; // outer_size * channels * channel_size
};
layout(binding = 1) buffer buf1{
float max_buffer[]; // outer_size * channel_size
};
layout(binding = 2) buffer buf2{
float sum_buffer[]; // outer_size * channel_size
};
layout(binding = 3) buffer buf3{
float output_buffer[]; // outer_size * channels * channel_size
};
layout(push_constant) uniform pushBlock {
int channel_size;
int outer_size;
int channels;
} p;
layout(local_size_x = LOCAL_SZ_X, local_size_y = 1, local_size_z = 1) in;
void main()
{
int gid = int(gl_GlobalInvocationID.x);
if (gid >= p.outer_size) return;
int global_off = gid * p.channels * p.channel_size;
int reduced_buffer_off = gid * p.channel_size;
// find the max along channel
int index = global_off;
for (int i = 0; i < p.channel_size; ++i)
{
max_buffer[reduced_buffer_off + i] = input_buffer[index];
index++;
}
for (int c = 1; c < p.channels; ++c)
{
for (int i = 0; i < p.channel_size; ++i)
{
max_buffer[reduced_buffer_off + i] = max(max_buffer[reduced_buffer_off + i], input_buffer[index]);
index++;
}
}
// substract, exp and accumulate along channel
for (int i = 0; i < p.channel_size; ++i)
sum_buffer[reduced_buffer_off + i] = 0.f;
index = global_off;
for (int c = 0; c < p.channels; ++c)
{
for (int i = 0; i < p.channel_size; ++i)
{
float exp_val = exp(input_buffer[index] - max_buffer[reduced_buffer_off + i]);
output_buffer[index] = exp_val;
sum_buffer[reduced_buffer_off + i] += exp_val;
index++;
}
}
// divide by computed sum
index = global_off;
for (int c = 0; c < p.channels; ++c)
{
for (int i = 0; i < p.channel_size; ++i)
{
float v = output_buffer[index] / sum_buffer[reduced_buffer_off + i];
#ifdef LOG_SOFTMAX
v = log(v);
#endif
output_buffer[index] = v;
index++;
}
}
}