COMPMID-927: Adding support for FP16 in CLDepthwiseConvolutionLayer3x3
Change-Id: Ie5f299c7a7fbe3062cee22bb2b4ae5df818fe490
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/121178
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
diff --git a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.h b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.h
index eb62465..fb51781 100644
--- a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.h
+++ b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.h
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -47,7 +47,7 @@
CLDepthwiseConvolutionLayer3x3Kernel &operator=(CLDepthwiseConvolutionLayer3x3Kernel &&) = default;
/** Initialize the function's source, destination, conv and border_size.
*
- * @param[in] input Source tensor. DataType supported: QASYMM8/F32.
+ * @param[in] input Source tensor. DataType supported: QASYMM8/F16/F32.
* @param[in] weights Weights tensor. A 3D tensor with dimensions [3, 3, IFM]. Data type supported: Same as @p input.
* @param[in] biases (Optional) Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
* Data type supported: Same as @p input.
diff --git a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
index b34967d..c348e3a 100644
--- a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
@@ -52,7 +52,7 @@
CLDepthwiseConvolutionLayer3x3();
/** Initialize the function's source, destination, conv and border_size.
*
- * @param[in, out] input Source tensor. Data type supported: QASYMM8/F32. (Written to only for border filling).
+ * @param[in, out] input Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
* @param[in] weights Weights tensor. A 3D tensor with shape [3, 3, IFM]. Data type supported: Same as @p input.
* @param[in] biases (Optional) Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
* Data type supported: Same as @p input.
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 5452b8a..c7c08d4 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -189,6 +189,7 @@
{ "copy_to_keypoint", "fast_corners.cl" },
{ "deconvolution_upsample", "deconvolution_layer.cl" },
{ "depthwise_convolution_3x3", "depthwise_convolution.cl" },
+ { "depthwise_convolution_3x3_f16", "depthwise_convolution.cl" },
{ "depthwise_convolution_3x3_quantized", "depthwise_convolution_quantized.cl" },
{ "depthwise_convolution_3x3_stridex1_stridey1_bifrost", "depthwise_convolution.cl" },
{ "depthwise_convolution_3x3_stridex2_stridey2_bifrost", "depthwise_convolution.cl" },
diff --git a/src/core/CL/cl_kernels/depthwise_convolution.cl b/src/core/CL/cl_kernels/depthwise_convolution.cl
index 8617886..f352138 100644
--- a/src/core/CL/cl_kernels/depthwise_convolution.cl
+++ b/src/core/CL/cl_kernels/depthwise_convolution.cl
@@ -564,3 +564,187 @@
}
#endif //defined(CONV_WIDTH) && defined(CONV_HEIGHT) && defined(DATA_TYPE)
+
+#if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED)
+#if defined(CONV_STRIDE_X)
+#if CONV_STRIDE_X == 1
+#define convolution1x3_f16 convolution1x3_stride_1_f16
+#elif CONV_STRIDE_X == 2
+#define convolution1x3_f16 convolution1x3_stride_2_f16
+#elif CONV_STRIDE_X == 3
+#define convolution1x3_f16 convolution1x3_stride_3_f16
+#else /* CONV_STRIDE_X */
+#error "Stride not supported"
+#endif /* CONV_STRIDE_X */
+
+/** Compute a 1D horizontal convolution of size 3 and stride 1 for 16bit floating point type.
+ *
+ * @param[in] left_pixel Pointer to the left pixel.
+ * @param[in] left_coeff Weight of the left pixel
+ * @param[in] middle_coeff Weight of the middle pixel
+ * @param[in] right_coeff Weight of the right pixel
+ *
+ * @return a half4 containing 4 convoluted values.
+ */
+inline half4 convolution1x3_stride_1_f16(__global const uchar *left_pixel,
+ const half left_coeff,
+ const half middle_coeff,
+ const half right_coeff)
+{
+ half8 temp = vload8(0, (__global half *)left_pixel);
+
+ half4 left = CONVERT(temp.s0123, half4);
+ half4 middle = CONVERT(temp.s1234, half4);
+ half4 right = CONVERT(temp.s2345, half4);
+
+ return left * (half4)left_coeff + middle * (half4)middle_coeff + right * (half4)right_coeff;
+}
+
+/** Compute a 1D horizontal convolution of size 3 and stride 2 for 16bit floating point type.
+ *
+ * @param[in] left_pixel Pointer to the left pixel.
+ * @param[in] left_coeff Weight of the left pixel
+ * @param[in] middle_coeff Weight of the middle pixel
+ * @param[in] right_coeff Weight of the right pixel
+ *
+ * @return a half4 containing 4 convoluted values.
+ */
+inline half4 convolution1x3_stride_2_f16(__global const uchar *left_pixel,
+ const half left_coeff,
+ const half middle_coeff,
+ const half right_coeff)
+{
+ half8 temp0 = vload8(0, (__global half *)left_pixel);
+ half temp1 = *((__global half *)(left_pixel + 8 * sizeof(half)));
+
+ half4 left = CONVERT(temp0.s0246, half4);
+ half4 middle = CONVERT(temp0.s1357, half4);
+ half4 right = CONVERT((half4)(temp0.s246, temp1), half4);
+
+ return left * (half4)left_coeff + middle * (half4)middle_coeff + right * (half4)right_coeff;
+}
+
+/** Compute a 1D horizontal convolution of size 3 and stride 3 for 16bit floating point type.
+ *
+ * @param[in] left_pixel Pointer to the left pixel.
+ * @param[in] left_coeff Weight of the left pixel
+ * @param[in] middle_coeff Weight of the middle pixel
+ * @param[in] right_coeff Weight of the right pixel
+ *
+ * @return a half4 containing 4 convoluted values.
+ */
+inline half4 convolution1x3_stride_3_f16(__global const uchar *left_pixel,
+ const half left_coeff,
+ const half middle_coeff,
+ const half right_coeff)
+{
+ half16 temp0 = vload16(0, (__global half *)left_pixel);
+
+ half4 left = CONVERT(temp0.s0369, half4);
+ half4 middle = CONVERT(temp0.s147A, half4);
+ half4 right = CONVERT(temp0.s258B, half4);
+
+ return left * (half4)left_coeff + middle * (half4)middle_coeff + right * (half4)right_coeff;
+}
+
+/** Apply a 3x3 convolution matrix to a single channel F16 input image and return the result.
+ *
+ * Convolution matrix layout:
+ *
+ * [ mat0, mat1, mat2 ]\n
+ * [ mat3, mat4, mat5 ]\n
+ * [ mat6, mat7, mat8 ]\n
+ *
+ * @param[in] src A pointer to source Image structure
+ * @param[in] mat0 Coefficient from the convolution matrix
+ * @param[in] mat1 Coefficient from the convolution matrix
+ * @param[in] mat2 Coefficient from the convolution matrix
+ * @param[in] mat3 Coefficient from the convolution matrix
+ * @param[in] mat4 Coefficient from the convolution matrix
+ * @param[in] mat5 Coefficient from the convolution matrix
+ * @param[in] mat6 Coefficient from the convolution matrix
+ * @param[in] mat0 Coefficient from the convolution matrix
+ * @param[in] mat7 Coefficient from the convolution matrix
+ * @param[in] mat8 Coefficient from the convolution matrix
+ *
+ * @return a half4 containing 4 convoluted values.
+ */
+inline half4 convolution3x3_f16(
+ Image *src,
+ const half mat0, const half mat1, const half mat2,
+ const half mat3, const half mat4, const half mat5,
+ const half mat6, const half mat7, const half mat8)
+{
+ half4 pixels;
+
+ pixels = convolution1x3_f16(offset(src, 0, 0), mat0, mat1, mat2);
+ pixels += convolution1x3_f16(offset(src, 0, 1), mat3, mat4, mat5);
+ pixels += convolution1x3_f16(offset(src, 0, 2), mat6, mat7, mat8);
+
+ return pixels;
+}
+
+/** This OpenCL kernel computes the depthwise convolution 3x3
+ *
+ * @param[in] src_ptr Pointer to the source image. Supported data types: F16
+ * @param[in] src_stride_x Stride of the source image in X dimension (in bytes)
+ * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] src_stride_y Stride of the source image in Y dimension (in bytes)
+ * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
+ * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: F32
+ * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
+ * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: F32
+ * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
+ * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
+ * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the biases vector
+ * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: F16/F32
+ * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes)
+ * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector
+ */
+__kernel void depthwise_convolution_3x3_f16(
+ TENSOR3D_DECLARATION(src),
+ TENSOR3D_DECLARATION(dst),
+ TENSOR3D_DECLARATION(weights)
+#if defined(HAS_BIAS)
+ ,
+ VECTOR_DECLARATION(biases)
+#endif //defined(HAS_BIAS)
+)
+{
+ Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src);
+ Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
+ Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT(weights);
+#if defined(HAS_BIAS)
+ Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
+#endif //defined(HAS_BIAS)
+
+ uchar3 offset = (uchar3)(0, 1, 2) * (uchar3)weights_stride_y;
+ half3 weights_values0 = vload3(0, (__global half *)(weights.ptr + offset.s0));
+ half3 weights_values1 = vload3(0, (__global half *)(weights.ptr + offset.s1));
+ half3 weights_values2 = vload3(0, (__global half *)(weights.ptr + offset.s2));
+
+ half4 pixels = convolution3x3_f16(&src, weights_values0.s0, weights_values0.s1, weights_values0.s2,
+ weights_values1.s0, weights_values1.s1, weights_values1.s2,
+ weights_values2.s0, weights_values2.s1, weights_values2.s2);
+#if defined(HAS_BIAS)
+ pixels += (half4)(*((__global half *)(biases.ptr + get_global_id(2) * biases_stride_x)));
+#endif //defined(HAS_BIAS)
+
+ vstore4(pixels, 0, (__global half *)dst.ptr);
+}
+#endif // defined(CONV_STRIDE_X)
+#endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED)
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.cpp
index 3613419..c7cee4c 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3Kernel.cpp
@@ -51,7 +51,7 @@
void CLDepthwiseConvolutionLayer3x3Kernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info)
{
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F32);
+ ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) != 3 || weights->info()->dimension(1) != 3);
@@ -134,7 +134,15 @@
// Create kernel
std::string kernel_name;
- if(input->info()->data_type() == DataType::F32 && gpu_target == GPUTarget::BIFROST)
+ if(input->info()->data_type() == DataType::F16)
+ {
+ kernel_name = "depthwise_convolution_3x3_f16";
+ num_elems_written_per_iteration_x = 8 / data_size_from_type(input->info()->data_type());
+ num_elems_written_per_iteration_y = 1;
+ num_elems_read_per_iteration_x = 3 + (num_elems_written_per_iteration_x - 1) * _conv_stride_x;
+ num_elems_read_per_iteration_y = 3;
+ }
+ else if(input->info()->data_type() == DataType::F32 && gpu_target == GPUTarget::BIFROST)
{
if(_conv_stride_x == 1 && _conv_stride_y == 1)
{
diff --git a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
index 84bfb94..88e9376 100644
--- a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
@@ -41,7 +41,7 @@
void CLDepthwiseConvolutionLayer3x3::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info)
{
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F32);
+ ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
_kernel.set_target(CLScheduler::get().target());
diff --git a/tests/validation/CL/DepthwiseConvolutionLayer.cpp b/tests/validation/CL/DepthwiseConvolutionLayer.cpp
index d747d43..20bf6cd 100644
--- a/tests/validation/CL/DepthwiseConvolutionLayer.cpp
+++ b/tests/validation/CL/DepthwiseConvolutionLayer.cpp
@@ -42,8 +42,9 @@
{
namespace
{
-constexpr RelativeTolerance<float> tolerance_f32(0.01f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
-constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); /**< Tolerance value for comparing reference's output against implementation's output for DataType::QASYMM8 */
+RelativeTolerance<half_float::half> tolerance_f16(half_float::half(0.001)); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */
+constexpr RelativeTolerance<float> tolerance_f32(0.01f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
+constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); /**< Tolerance value for comparing reference's output against implementation's output for DataType::QASYMM8 */
} // namespace
TEST_SUITE(CL)
@@ -70,6 +71,23 @@
using CLDepthwiseConvolutionLayerFixture3x3 = DepthwiseConvolutionLayerValidationFixture<CLTensor, CLAccessor, CLDepthwiseConvolutionLayer3x3, T>;
TEST_SUITE(Float)
+TEST_SUITE(F16)
+TEST_SUITE(W3x3)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerFixture3x3<half>, framework::DatasetMode::ALL, combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(),
+ framework::dataset::make("DataType",
+ DataType::F16)))
+{
+ validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerFixture3x3<half>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(),
+ framework::dataset::make("DataType",
+ DataType::F16)))
+{
+ validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+TEST_SUITE_END()
+TEST_SUITE_END()
+
TEST_SUITE(FP32)
TEST_SUITE(W3x3)
FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerFixture3x3<float>, framework::DatasetMode::ALL, combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(),