From 9f7d55a3566b0f1044110000b033d663b26d3a6c Mon Sep 17 00:00:00 2001 From: Giorgio Arena Date: Mon, 8 Feb 2021 13:20:24 +0000 Subject: Fix CLDepthwiseConvolutionLayer 3x3 QASYMM8 Fix errors when computing tensors with one element only - Replace Tensor3D with raw pointers so to get rid of offset to first element for NCHW layout - Add stronger out of bound constraints for NHWC layout - Set the border size to the input's padding for NHWC - Fill the strides == 0 with the largest stride, so to avoid accessing empty strides and multiplying by 0 Resolve COMPMID-4088 Change-Id: I751a4e6d7094b3c42306ff7f53af848fd35f19ac Signed-off-by: Giorgio Arena Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5024 Tested-by: Arm Jenkins Reviewed-by: Manuel Bottini Reviewed-by: Michele Di Giorgio Comments-Addressed: Arm Jenkins --- .../cl_kernels/depthwise_convolution_quantized.cl | 44 ++++++++++++---------- 1 file changed, 24 insertions(+), 20 deletions(-) (limited to 'src/core/CL/cl_kernels/depthwise_convolution_quantized.cl') diff --git a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl index d39089b923..285c00a713 100644 --- a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl +++ b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl @@ -194,11 +194,11 @@ __kernel void dwc_3x3_native_quantized8_nchw( #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_NO_STEP(weights); - Vector output_multipliers = CONVERT_TO_VECTOR_STRUCT_NO_STEP(output_multipliers); - Vector output_shifts = CONVERT_TO_VECTOR_STRUCT_NO_STEP(output_shifts); + __global uchar *src_addr = src_ptr + get_global_id(0) * src_step_x + get_global_id(1) * src_step_y + get_global_id(2) * src_step_z; + Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst); + Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); + Vector output_multipliers = CONVERT_TO_VECTOR_STRUCT_NO_STEP(output_multipliers); + Vector output_shifts = CONVERT_TO_VECTOR_STRUCT_NO_STEP(output_shifts); // Extract channel and linearized batch indices const int channel = get_global_id(2) % DST_CHANNELS; @@ -211,7 +211,7 @@ __kernel void dwc_3x3_native_quantized8_nchw( #endif //defined(HAS_BIAS) // Load relevant input and weights data (Accounts depth multiplier when indexing input, OFM = IFM * DEPTH_MULTIPLIER) - src.ptr -= batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z + (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z; + src_addr -= batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z + (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z; __global uchar *weights_addr = weights.ptr + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z; VEC_DATA_TYPE(WEIGHTS_TYPE, 3) @@ -235,7 +235,7 @@ __kernel void dwc_3x3_native_quantized8_nchw( // Row0 int8 left, middle, right; - GET_VALUES(src.ptr + 0 * src_stride_y, left, middle, right); + GET_VALUES(src_addr + 0 * src_stride_y, left, middle, right); values0 += left * (int8)(w0.s0); values0 += middle * (int8)(w0.s1); values0 += right * (int8)(w0.s2); @@ -245,10 +245,11 @@ __kernel void dwc_3x3_native_quantized8_nchw( #endif /* WEIGHTS_OFFSET != 0 */ // Row1 - GET_VALUES(src.ptr + DILATION_Y * src_stride_y, left, middle, right); + GET_VALUES(src_addr + DILATION_Y * src_stride_y, left, middle, right); values0 += left * (int8)(w1.s0); values0 += middle * (int8)(w1.s1); values0 += right * (int8)(w1.s2); + #if CONV_STRIDE_Y == 1 && DILATION_Y == 1 values1 += left * (int8)(w0.s0); values1 += middle * (int8)(w0.s1); @@ -264,7 +265,7 @@ __kernel void dwc_3x3_native_quantized8_nchw( #endif /* WEIGHTS_OFFSET != 0 */ // Row2 - GET_VALUES(src.ptr + 2 * DILATION_Y * src_stride_y, left, middle, right); + GET_VALUES(src_addr + 2 * DILATION_Y * src_stride_y, left, middle, right); values0 += left * (int8)(w2.s0); values0 += middle * (int8)(w2.s1); values0 += right * (int8)(w2.s2); @@ -284,7 +285,7 @@ __kernel void dwc_3x3_native_quantized8_nchw( #if CONV_STRIDE_Y == 1 && DILATION_Y == 1 // Row3 - GET_VALUES(src.ptr + 3 * src_stride_y, left, middle, right); + GET_VALUES(src_addr + 3 * src_stride_y, left, middle, right); values1 += left * (int8)(w2.s0); values1 += middle * (int8)(w2.s1); values1 += right * (int8)(w2.s2); @@ -511,11 +512,11 @@ __kernel void dwc_3x3_native_quantized8_dot8_nchw( #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_NO_STEP(weights); - Vector output_multipliers = CONVERT_TO_VECTOR_STRUCT_NO_STEP(output_multipliers); - Vector output_shifts = CONVERT_TO_VECTOR_STRUCT_NO_STEP(output_shifts); + __global uchar *src_addr = src_ptr + get_global_id(0) * src_step_x + get_global_id(1) * src_step_y + get_global_id(2) * src_step_z; + Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst); + Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); + Vector output_multipliers = CONVERT_TO_VECTOR_STRUCT_NO_STEP(output_multipliers); + Vector output_shifts = CONVERT_TO_VECTOR_STRUCT_NO_STEP(output_shifts); // Extract channel and linearized batch indices const int channel = get_global_id(2) % DST_CHANNELS; @@ -528,7 +529,7 @@ __kernel void dwc_3x3_native_quantized8_dot8_nchw( #endif //defined(HAS_BIAS) // Load relevant input and weights data (Accounts depth multiplier when indexing input, OFM = IFM * DEPTH_MULTIPLIER) - src.ptr -= batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z + (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z; + src_addr -= batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z + (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z; __global uchar *weights_addr = weights.ptr + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z; VEC_TYPE(3) @@ -551,9 +552,9 @@ __kernel void dwc_3x3_native_quantized8_dot8_nchw( int8 values0 = 0; int8 sum0 = 0; - GET_VALUES(src.ptr + 0 * src_stride_y, left0, middle0, right0); - GET_VALUES(src.ptr + DILATION_Y * src_stride_y, left1, middle1, right1); - GET_VALUES(src.ptr + 2 * DILATION_Y * src_stride_y, left2, middle2, right2); + GET_VALUES(src_addr + 0 * src_stride_y, left0, middle0, right0); + GET_VALUES(src_addr + DILATION_Y * src_stride_y, left1, middle1, right1); + GET_VALUES(src_addr + 2 * DILATION_Y * src_stride_y, left2, middle2, right2); #if WEIGHTS_OFFSET != 0 sum0 += convert_int8(left0) + convert_int8(middle0) + convert_int8(right0); @@ -569,7 +570,7 @@ __kernel void dwc_3x3_native_quantized8_dot8_nchw( int8 values1 = 0; int8 sum1 = 0; - GET_VALUES(src.ptr + 3 * src_stride_y, left3, middle3, right3); + GET_VALUES(src_addr + 3 * src_stride_y, left3, middle3, right3); #if WEIGHTS_OFFSET != 0 sum1 += convert_int8(left1) + convert_int8(middle1) + convert_int8(right1); @@ -923,7 +924,9 @@ __kernel void dwc_3x3_reshaped_quantized8_nhwc( // z_coord can be only negative for z = 0 so we do not need to clamp it // Moreover z_coord cannot be out-of-bound for z = 1 so we do not need to clamp the offset z_coord = z * (int)CONV_STRIDE_Y - (int)CONV_PAD_TOP + DILATION_Y; + z_coord = min((uint)z_coord, (uint)SRC_DIM_2); offset = y_offset + (int4)(z_coord * src_stride_z); + offset = min(offset, (int4)max_offset); VEC_TYPE(VEC_SIZE) values3 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(src_addr + offset.s0)); VEC_TYPE(VEC_SIZE) @@ -934,6 +937,7 @@ __kernel void dwc_3x3_reshaped_quantized8_nhwc( // z == 2 // Offset can be out-of-bound so we need to check if it is greater than max_offset z_coord = z * (int)CONV_STRIDE_Y - (int)CONV_PAD_TOP + DILATION_Y * 2; + z_coord = min((uint)z_coord, (uint)SRC_DIM_2); offset = y_offset + (int4)(z_coord * src_stride_z); offset = min(offset, (int4)max_offset); VEC_TYPE(VEC_SIZE) -- cgit v1.2.1