From e5362e7e5dbccf81c5296a7e77154e11e1a14d2f Mon Sep 17 00:00:00 2001 From: Pablo Marquez Tello Date: Mon, 16 Oct 2023 14:14:14 +0100 Subject: DirectConv and Im2Col changes to enable fp16 in armv8a multi_isa builds * FP16 kernels must be instantiated in fp16.cpp. * Partially resolves MLCE-1102 Change-Id: Iab9c29dbfd89358f2f663862ff5010c88aeccf8c Signed-off-by: Pablo Marquez Tello Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10496 Reviewed-by: Anitha Raj Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Benchmark: Arm Jenkins --- src/cpu/kernels/CpuIm2ColKernel.cpp | 569 +++++++++++++----------------------- 1 file changed, 210 insertions(+), 359 deletions(-) (limited to 'src/cpu/kernels/CpuIm2ColKernel.cpp') diff --git a/src/cpu/kernels/CpuIm2ColKernel.cpp b/src/cpu/kernels/CpuIm2ColKernel.cpp index 55ac7c5192..39ba764c78 100644 --- a/src/cpu/kernels/CpuIm2ColKernel.cpp +++ b/src/cpu/kernels/CpuIm2ColKernel.cpp @@ -35,6 +35,8 @@ #include "src/core/CPP/Validate.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" +#include "src/cpu/kernels/directconv2d/impl.h" +#include "src/cpu/kernels/directconv2d/list.h" #include #include @@ -49,6 +51,198 @@ namespace cpu { namespace kernels { +void run_im2col_fp32_pad(const ITensor *src, + ITensor *dst, + const Window &window, + DataLayout data_layout, + const PadStrideInfo &conv_info, + std::pair convolved_dims, + const Size2D &kernel_dims, + const Size2D &dilation, + uint32_t input_pad_right, + bool has_bias) +{ + arm_compute::cpu::kernels::run_im2col(src, dst, window, data_layout, conv_info, convolved_dims, + kernel_dims, dilation, input_pad_right, has_bias); +} + +void run_im2col_fp32_nopad(const ITensor *src, + ITensor *dst, + const Window &window, + DataLayout data_layout, + const PadStrideInfo &conv_info, + std::pair convolved_dims, + const Size2D &kernel_dims, + const Size2D &dilation, + uint32_t input_pad_right, + bool has_bias) +{ + arm_compute::cpu::kernels::run_im2col(src, dst, window, data_layout, conv_info, convolved_dims, + kernel_dims, dilation, input_pad_right, has_bias); +} + +#if defined(ARM_COMPUTE_ENABLE_BF16) +void run_im2col_bf16_pad(const ITensor *src, + ITensor *dst, + const Window &window, + DataLayout data_layout, + const PadStrideInfo &conv_info, + std::pair convolved_dims, + const Size2D &kernel_dims, + const Size2D &dilation, + uint32_t input_pad_right, + bool has_bias) +{ + arm_compute::cpu::kernels::run_im2col( + src, dst, window, data_layout, conv_info, convolved_dims, kernel_dims, dilation, input_pad_right, has_bias); +} + +void run_im2col_bf16_nopad(const ITensor *src, + ITensor *dst, + const Window &window, + DataLayout data_layout, + const PadStrideInfo &conv_info, + std::pair convolved_dims, + const Size2D &kernel_dims, + const Size2D &dilation, + uint32_t input_pad_right, + bool has_bias) +{ + arm_compute::cpu::kernels::run_im2col( + src, dst, window, data_layout, conv_info, convolved_dims, kernel_dims, dilation, input_pad_right, has_bias); +} +#endif /* defined(ARM_COMPUTE_ENABLE_BF16) */ + +void run_im2col_int8_nopad_nhwc(const ITensor *src, + ITensor *dst, + const Window &window, + DataLayout data_layout, + const PadStrideInfo &conv_info, + std::pair convolved_dims, + const Size2D &kernel_dims, + const Size2D &dilation, + uint32_t input_pad_right, + bool has_bias) +{ + arm_compute::cpu::kernels::run_im2col( + src, dst, window, data_layout, conv_info, convolved_dims, kernel_dims, dilation, input_pad_right, has_bias); +} + +void run_im2col_uint8_nopad_nhwc(const ITensor *src, + ITensor *dst, + const Window &window, + DataLayout data_layout, + const PadStrideInfo &conv_info, + std::pair convolved_dims, + const Size2D &kernel_dims, + const Size2D &dilation, + uint32_t input_pad_right, + bool has_bias) +{ + arm_compute::cpu::kernels::run_im2col( + src, dst, window, data_layout, conv_info, convolved_dims, kernel_dims, dilation, input_pad_right, has_bias); +} + +void run_im2col_qasymm8_pad_nhwc(const ITensor *src, + ITensor *dst, + const Window &window, + DataLayout data_layout, + const PadStrideInfo &conv_info, + std::pair convolved_dims, + const Size2D &kernel_dims, + const Size2D &dilation, + uint32_t input_pad_right, + bool has_bias) +{ + arm_compute::cpu::kernels::run_im2col( + src, dst, window, data_layout, conv_info, convolved_dims, kernel_dims, dilation, input_pad_right, has_bias); +} + +void internal_run_im2col_fp16_pad(const ITensor *src, + ITensor *dst, + const Window &window, + DataLayout data_layout, + const PadStrideInfo &conv_info, + std::pair convolved_dims, + const Size2D &kernel_dims, + const Size2D &dilation, + uint32_t input_pad_right, + bool has_bias) +{ +/* + Note that when building with the option data_type_support=fp32 the fp16.cpp files won't be compiled and the linker + would fail with the error undefined arm_compute::cpu::kernels::run_im2col_fp16_pad. + To avoid this problem we only call to the actual fp16 kernel if ENABLE_FP16_KERNELS is defined. +*/ +#if defined(ENABLE_FP16_KERNELS) + arm_compute::cpu::kernels::run_im2col_fp16_pad(src, dst, window, data_layout, conv_info, convolved_dims, + kernel_dims, dilation, input_pad_right, has_bias); +#else // defined(ENABLE_FP16_KERNELS) + ARM_COMPUTE_UNUSED(src, dst, window, data_layout, conv_info, convolved_dims, kernel_dims, dilation, input_pad_right, + has_bias); +#endif // defined(ENABLE_FP16_KERNELS) +} + +void internal_run_im2col_fp16_nopad(const ITensor *src, + ITensor *dst, + const Window &window, + DataLayout data_layout, + const PadStrideInfo &conv_info, + std::pair convolved_dims, + const Size2D &kernel_dims, + const Size2D &dilation, + uint32_t input_pad_right, + bool has_bias) +{ +#if defined(ENABLE_FP16_KERNELS) + arm_compute::cpu::kernels::run_im2col_fp16_nopad(src, dst, window, data_layout, conv_info, convolved_dims, + kernel_dims, dilation, input_pad_right, has_bias); +#else // defined(ENABLE_FP16_KERNELS) + ARM_COMPUTE_UNUSED(src, dst, window, data_layout, conv_info, convolved_dims, kernel_dims, dilation, input_pad_right, + has_bias); +#endif // defined(ENABLE_FP16_KERNELS) +} + +void internal_run_im2col_fp16_nchw_pad(const ITensor *src, + ITensor *dst, + const Window &window, + DataLayout data_layout, + const PadStrideInfo &conv_info, + std::pair convolved_dims, + const Size2D &kernel_dims, + const Size2D &dilation, + uint32_t input_pad_right, + bool has_bias) +{ +#if defined(ENABLE_FP16_KERNELS) + arm_compute::cpu::kernels::run_im2col_fp16_nchw_pad(src, dst, window, data_layout, conv_info, convolved_dims, + kernel_dims, dilation, input_pad_right, has_bias); +#else // defined(ENABLE_FP16_KERNELS) + ARM_COMPUTE_UNUSED(src, dst, window, data_layout, conv_info, convolved_dims, kernel_dims, dilation, input_pad_right, + has_bias); +#endif // defined(ENABLE_FP16_KERNELS) +} + +void internal_run_im2col_fp16_nchw_nopad(const ITensor *src, + ITensor *dst, + const Window &window, + DataLayout data_layout, + const PadStrideInfo &conv_info, + std::pair convolved_dims, + const Size2D &kernel_dims, + const Size2D &dilation, + uint32_t input_pad_right, + bool has_bias) +{ +#if defined(ENABLE_FP16_KERNELS) + arm_compute::cpu::kernels::run_im2col_fp16_nchw_nopad(src, dst, window, data_layout, conv_info, convolved_dims, + kernel_dims, dilation, input_pad_right, has_bias); +#else // defined(ENABLE_FP16_KERNELS) + ARM_COMPUTE_UNUSED(src, dst, window, data_layout, conv_info, convolved_dims, kernel_dims, dilation, input_pad_right, + has_bias); +#endif // defined(ENABLE_FP16_KERNELS) +} + namespace { Status validate_arguments(const ITensorInfo *input, @@ -86,340 +280,8 @@ Status validate_arguments(const ITensorInfo *input, return Status{}; } - -template -inline void linearize_volume_nchw(const uint8_t *const in_ptr, - T *out_ptr, - bool has_bias, - int top_left_x, - int top_left_y, - int kernel_width, - int kernel_height, - int kernel_depth, - int input_w, - int input_h, - int input_stride_x, - int input_stride_y, - int input_stride_z, - int pad_value, - int dilation_x, - int dilation_y) -{ - const int kernel_size2 = kernel_width * kernel_height; - const int x_e = top_left_x + kernel_width * dilation_x; - const int y_e = top_left_y + kernel_height * dilation_y; - - // Linearize volume - int d = 0; - // This for loop linearize a volume with 3 slices. This allows: - // 1) to reduce the iterations of the outer for loop "d" - // 2) to have an optimized im2col for the first convolution layer where usually we have 3 IFMs - for (; d <= (kernel_depth - 3); d += 3) - { - for (int y = top_left_y; y < y_e; y += dilation_y) - { - if ((y < 0 || y >= input_h) && has_pads) - { - // All the values will be the offset (will be zeros when not quantized) - for (int x = top_left_x; x < x_e; x += dilation_x, ++out_ptr) - { - *(out_ptr + 0 * kernel_size2) = pad_value; - *(out_ptr + 1 * kernel_size2) = pad_value; - *(out_ptr + 2 * kernel_size2) = pad_value; - } - } - else - { - for (int x = top_left_x; x < x_e; x += dilation_x, ++out_ptr) - { - if ((x < 0 || x >= input_w) && has_pads) - { - *(out_ptr + 0 * kernel_size2) = pad_value; - *(out_ptr + 1 * kernel_size2) = pad_value; - *(out_ptr + 2 * kernel_size2) = pad_value; - } - else - { - *(out_ptr + 0 * kernel_size2) = *(reinterpret_cast( - in_ptr + ((d + 0) * input_stride_z + y * input_stride_y + x * input_stride_x))); - *(out_ptr + 1 * kernel_size2) = *(reinterpret_cast( - in_ptr + ((d + 1) * input_stride_z + y * input_stride_y + x * input_stride_x))); - *(out_ptr + 2 * kernel_size2) = *(reinterpret_cast( - in_ptr + ((d + 2) * input_stride_z + y * input_stride_y + x * input_stride_x))); - } - } - } - } - out_ptr += 2 * kernel_size2; - } - - // Left over - for (; d < kernel_depth; d++) - { - for (int y = top_left_y; y < y_e; y += dilation_y) - { - if ((y < 0 || y >= input_h) && has_pads) - { - // All the values will be the offset (will be zeros when not quantized) - memset(static_cast(out_ptr), pad_value, kernel_width * sizeof(T)); - out_ptr += kernel_width; - } - else - { - for (int x = top_left_x; x < x_e; x += dilation_x, ++out_ptr) - { - if ((x < 0 || x >= input_w) && has_pads) - { - *out_ptr = pad_value; - } - else - { - *out_ptr = *(reinterpret_cast( - in_ptr + (d * input_stride_z + y * input_stride_y + x * input_stride_x))); - } - } - } - } - } - - // Append 1 if the convolution layer has biases - if (has_bias) - { - *out_ptr = static_cast(1); - } -} - -template -inline void linearize_volume_nhwc(const uint8_t *const in_ptr, - T *out_ptr, - bool has_bias, - int start_x, - int start_y, - int kernel_width, - int kernel_height, - int input_w, - int input_h, - int input_c, - int input_stride_y, - int input_stride_z, - int pad_value, - int dilation_x, - int dilation_y) -{ - const int end_x = start_x + kernel_width * dilation_x; - const int end_y = start_y + kernel_height * dilation_y; - const int pad_quant = kernel_width * input_c; - const int element_size = static_cast(sizeof(T)); - if ((start_y >= 0) && (end_y < input_h) && (start_x >= 0) && (end_x < input_w) && (dilation_x == 1) && - (input_stride_y == input_c * element_size)) - { - for (int y = start_y; y < end_y; y += dilation_y) - { - //optimized for no dilation and no boundary pixels - memcpy(out_ptr, reinterpret_cast(in_ptr + (y * input_stride_z + start_x * input_stride_y)), - input_c * kernel_width * element_size); - out_ptr += input_c * kernel_width; - } - } - else - { - for (int y = start_y; y < end_y; y += dilation_y) - { - if (y < 0 || y >= input_h) - { - memset(static_cast(out_ptr), pad_value, pad_quant * element_size); - out_ptr += pad_quant; - } - else if (dilation_x > 1 || start_x < 0 || end_x >= input_w || input_stride_y != input_c * element_size) - { - for (int x = start_x; x < end_x; x += dilation_x) - { - if (x < 0 || x >= input_w) - { - memset(static_cast(out_ptr), pad_value, input_c * element_size); - out_ptr += input_c; - } - else - { - memcpy(out_ptr, reinterpret_cast(in_ptr + (y * input_stride_z + x * input_stride_y)), - input_c * element_size); - out_ptr += input_c; - } - } - } - else - { - //optimized for no dilation and no boundary pixels - memcpy(out_ptr, reinterpret_cast(in_ptr + (y * input_stride_z + start_x * input_stride_y)), - input_c * kernel_width * element_size); - out_ptr += input_c * kernel_width; - } - } - } - // Append 1 if the convolution layer has biases - if (has_bias) - { - *out_ptr = static_cast(1); - } -} - -template -inline void linearize_volume_nhwc(const uint8_t *const in_ptr, - T *out_ptr, - bool has_bias, - int start_x, - int start_y, - int kernel_width, - int kernel_height, - int input_w, - int input_h, - int input_c, - int input_stride_y, - int input_stride_z, - int pad_value, - int dilation_x, - int dilation_y, - int pad_right) -{ - const int end_x = start_x + kernel_width * dilation_x; - const int end_y = start_y + kernel_height * dilation_y; - const int pad_quant = kernel_width * (input_c + pad_right); - const int element_size = static_cast(sizeof(T)); - const int channel_chunk_size = input_c * element_size; - - if ((start_y >= 0) && (end_y < input_h) && (start_x >= 0) && (end_x < input_w) && (dilation_x == 1) && - (input_stride_y == channel_chunk_size)) - { - for (int y = start_y; y < end_y; y += dilation_y) - { - const uint8_t *offset_ptr = in_ptr + (y * input_stride_z + start_x * input_stride_y); - for (int e = 0; e < kernel_width; e++) - { - memcpy(out_ptr, reinterpret_cast(offset_ptr + e * channel_chunk_size), channel_chunk_size); - out_ptr += input_c + pad_right; - } - } - } - else - { - for (int y = start_y; y < end_y; y += dilation_y) - { - if (y < 0 || y >= input_h) - { - memset(static_cast(out_ptr), pad_value, pad_quant * element_size); - out_ptr += pad_quant; - } - else if (dilation_x > 1 || start_x < 0 || end_x >= input_w || input_stride_y != channel_chunk_size) - { - for (int x = start_x; x < end_x; x += dilation_x) - { - if (x < 0 || x >= input_w) - { - memset(static_cast(out_ptr), pad_value, (input_c + pad_right) * element_size); - out_ptr += input_c + pad_right; - } - else - { - memcpy(out_ptr, reinterpret_cast(in_ptr + (y * input_stride_z + x * input_stride_y)), - channel_chunk_size); - out_ptr += input_c + pad_right; - } - } - } - else - { - const uint8_t *offset_ptr = in_ptr + (y * input_stride_z + start_x * input_stride_y); - for (int e = 0; e < kernel_width; e++) - { - memcpy(out_ptr, reinterpret_cast(offset_ptr + e * channel_chunk_size), - channel_chunk_size); - out_ptr += input_c + pad_right; - } - } - } - } - // Append 1 if the convolution layer has biases - if (has_bias) - { - *out_ptr = static_cast(1); - } -} - } // namespace -template -void CpuIm2ColKernel::run_im2col(const ITensor *src, ITensor *dst, const Window &window) -{ - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window); - - const unsigned int width_idx = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH); - const unsigned int height_idx = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT); - const unsigned int channel_idx = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL); - - const int input_w = src->info()->dimension(width_idx); - const int input_h = src->info()->dimension(height_idx); - const int input_c = src->info()->dimension(channel_idx); - const int input_stride_x = src->info()->strides_in_bytes().x(); - const int input_stride_y = src->info()->strides_in_bytes().y(); - const int input_stride_z = src->info()->strides_in_bytes().z(); - const int pad_left = _conv_info.pad_left(); - const int pad_top = _conv_info.pad_top(); - const int stride_x = _conv_info.stride().first; - const int stride_y = _conv_info.stride().second; - const int pad_value = - is_data_type_quantized(src->info()->data_type()) ? src->info()->quantization_info().uniform().offset : 0; - - Window window_in_out(window); - // The first three dimensions of the input and output are increased by the inner loops - window_in_out.set(Window::DimX, Window::Dimension(0, 0, 0)); - window_in_out.set(Window::DimY, Window::Dimension(0, 0, 0)); - window_in_out.set(Window::DimZ, Window::Dimension(0, 0, 0)); - - // Create iterators - Iterator in(src, window_in_out); - Iterator out(dst, window_in_out); - - execute_window_loop( - window, - [&](const Coordinates &id) - { - const int start_w = id[width_idx] * stride_x - pad_left; - const int start_h = id[height_idx] * stride_y - pad_top; - - // Get pointers - const uint8_t *const input_ptr = in.ptr(); - auto output_ptr = - reinterpret_cast(out.ptr() + (id[width_idx] + id[height_idx] * _convolved_dims.first) * - dst->info()->strides_in_bytes().y()); - - // Linearize volume - if (is_nchw) - { - linearize_volume_nchw( - input_ptr, output_ptr, _has_bias, start_w, start_h, _kernel_width, _kernel_height, input_c, input_w, - input_h, input_stride_x, input_stride_y, input_stride_z, pad_value, _dilation.x(), _dilation.y()); - } - else - { - if (_input_pad_right > 0) - { - linearize_volume_nhwc(input_ptr, output_ptr, _has_bias, start_w, start_h, - _kernel_width, _kernel_height, input_w, input_h, input_c, - input_stride_y, input_stride_z, pad_value, _dilation.x(), - _dilation.y(), _input_pad_right); - } - else - { - linearize_volume_nhwc( - input_ptr, output_ptr, _has_bias, start_w, start_h, _kernel_width, _kernel_height, input_w, - input_h, input_c, input_stride_y, input_stride_z, pad_value, _dilation.x(), _dilation.y()); - } - } - }, - in, out); -} - void CpuIm2ColKernel::configure(const ITensorInfo *src, ITensorInfo *dst, const Size2D &kernel_dims, @@ -453,25 +315,20 @@ void CpuIm2ColKernel::configure(const ITensorInfo *src, switch (src->data_type()) { case DataType::F32: - _func = (!conv_info.has_padding()) ? &CpuIm2ColKernel::run_im2col - : &CpuIm2ColKernel::run_im2col; + _func = (!conv_info.has_padding()) ? &run_im2col_fp32_nchw_nopad : &run_im2col_fp32_nchw_pad; + break; + case DataType::F16: + _func = (!conv_info.has_padding()) ? &internal_run_im2col_fp16_nchw_nopad + : &internal_run_im2col_fp16_nchw_pad; break; #if defined(ARM_COMPUTE_ENABLE_BF16) case DataType::BFLOAT16: - _func = (!conv_info.has_padding()) ? &CpuIm2ColKernel::run_im2col - : &CpuIm2ColKernel::run_im2col; + _func = (!conv_info.has_padding()) ? &run_im2col_bf16_nchw_nopad : &run_im2col_bf16_nchw_pad; break; #endif /* defined(ARM_COMPUTE_ENABLE_BF16) */ -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - case DataType::F16: - _func = (!conv_info.has_padding()) ? &CpuIm2ColKernel::run_im2col - : &CpuIm2ColKernel::run_im2col; - break; -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ case DataType::QASYMM8_SIGNED: case DataType::QASYMM8: - _func = (!conv_info.has_padding()) ? &CpuIm2ColKernel::run_im2col - : &CpuIm2ColKernel::run_im2col; + _func = (!conv_info.has_padding()) ? &run_im2col_qasymm8_nchw_nopad : &run_im2col_qasymm8_nchw_pad; break; default: ARM_COMPUTE_ERROR("Data type not supported"); @@ -483,28 +340,21 @@ void CpuIm2ColKernel::configure(const ITensorInfo *src, switch (src->data_type()) { case DataType::F32: - _func = (!conv_info.has_padding()) ? &CpuIm2ColKernel::run_im2col - : &CpuIm2ColKernel::run_im2col; + _func = (!conv_info.has_padding()) ? &run_im2col_fp32_nopad : &run_im2col_fp32_pad; + break; + case DataType::F16: + _func = (!conv_info.has_padding()) ? &internal_run_im2col_fp16_nopad : &internal_run_im2col_fp16_pad; break; #if defined(ARM_COMPUTE_ENABLE_BF16) case DataType::BFLOAT16: - _func = (!conv_info.has_padding()) ? &CpuIm2ColKernel::run_im2col - : &CpuIm2ColKernel::run_im2col; + _func = (!conv_info.has_padding()) ? &run_im2col_bf16_nopad : &run_im2col_bf16_pad; break; #endif /* defined(ARM_COMPUTE_ENABLE_BF16) */ -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - case DataType::F16: - _func = (!conv_info.has_padding()) ? &CpuIm2ColKernel::run_im2col - : &CpuIm2ColKernel::run_im2col; - break; -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ case DataType::QASYMM8: - _func = (!conv_info.has_padding()) ? &CpuIm2ColKernel::run_im2col - : &CpuIm2ColKernel::run_im2col; + _func = (!conv_info.has_padding()) ? &run_im2col_uint8_nopad_nhwc : &run_im2col_qasymm8_pad_nhwc; break; case DataType::QASYMM8_SIGNED: - _func = (!conv_info.has_padding()) ? &CpuIm2ColKernel::run_im2col - : &CpuIm2ColKernel::run_im2col; + _func = (!conv_info.has_padding()) ? &run_im2col_int8_nopad_nhwc : &run_im2col_qasymm8_pad_nhwc; break; default: ARM_COMPUTE_ERROR("Data type not supported"); @@ -552,7 +402,8 @@ void CpuIm2ColKernel::run_op(ITensorPack &tensors, const Window &window, const T auto src = tensors.get_const_tensor(TensorType::ACL_SRC); auto dst = tensors.get_tensor(TensorType::ACL_DST); - (this->*_func)(src, dst, window); + _func(src, dst, window, _data_layout, _conv_info, _convolved_dims, Size2D(_kernel_width, _kernel_height), _dilation, + _input_pad_right, _has_bias); } const char *CpuIm2ColKernel::name() const -- cgit v1.2.1