Subpixel layout support for RGB/BGR to greyscale (ROCm#113)

* Mods to handle RGB and BGR subpixel layouts * Test suite mods to handle RGB/BGR subpixel layouts in color to greyscale conversions * Rename files from rgb_to_greyscale to color_to_greyscale
Akilesh2 · Jul 13, 2022 · 64dd326 · 64dd326
1 parent ca3bd97
commit 64dd326
Show file tree

Hide file tree

Showing 21 changed files with 440 additions and 371 deletions.
diff --git a/include/rppdefs.h b/include/rppdefs.h
@@ -220,9 +220,14 @@ typedef enum
 {
     LTRB,
     XYWH
-
 } RpptRoiType;
 
+typedef enum
+{
+    RGBtype,
+    BGRtype
+} RpptSubpixelLayout;
+
 typedef enum
 {
     NEAREST_NEIGHBOR = 0,

diff --git a/include/rppt_tensor_data_exchange_operations.h b/include/rppt_tensor_data_exchange_operations.h
@@ -62,21 +62,22 @@ RppStatus rppt_swap_channels_host(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppPt
 RppStatus rppt_swap_channels_gpu(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppPtr_t dstPtr, RpptDescPtr dstDescPtr, rppHandle_t rppHandle);
 #endif // GPU_SUPPORT
 
-/******************** rgb_to_greyscale ********************/
+/******************** color_to_greyscale ********************/
 
-// RGB to greyscale operation for a NCHW/NHWC layout tensor
+// Color to greyscale operation for a RGB/BGR NCHW/NHWC layout tensor
 
 // *param[in] srcPtr source tensor memory
 // *param[in] srcDescPtr source tensor descriptor
 // *param[out] dstPtr destination tensor memory
 // *param[in] dstDescPtr destination tensor descriptor
+// *param[in] srcSubpixelLayout A RpptSubpixelLayout type enum to specify source subpixel layout (RGBtype or BGRtype)
 // *returns a  RppStatus enumeration.
 // *retval RPP_SUCCESS : succesful completion
 // *retval RPP_ERROR : Error
 
-RppStatus rppt_rgb_to_greyscale_host(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppPtr_t dstPtr, RpptDescPtr dstDescPtr, rppHandle_t rppHandle);
+RppStatus rppt_color_to_greyscale_host(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppPtr_t dstPtr, RpptDescPtr dstDescPtr, RpptSubpixelLayout srcSubpixelLayout, rppHandle_t rppHandle);
 #ifdef GPU_SUPPORT
-RppStatus rppt_rgb_to_greyscale_gpu(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppPtr_t dstPtr, RpptDescPtr dstDescPtr, rppHandle_t rppHandle);
+RppStatus rppt_color_to_greyscale_gpu(RppPtr_t srcPtr, RpptDescPtr srcDescPtr, RppPtr_t dstPtr, RpptDescPtr dstDescPtr, RpptSubpixelLayout srcSubpixelLayout, rppHandle_t rppHandle);
 #endif // GPU_SUPPORT
 
 #ifdef __cplusplus

diff --git a/src/include/cpu/rpp_cpu_common.hpp b/src/include/cpu/rpp_cpu_common.hpp
@@ -2146,28 +2146,28 @@ inline RppStatus custom_convolve_image_host(T* srcPtr, RppiSize srcSize, U* dstP
 
 // Compute Functions for RPP Tensor API
 
-inline void compute_rgb_48_to_greyscale_16_host(__m256 *p, __m256 *pWeightsRGB)
+inline void compute_color_48_to_greyscale_16_host(__m256 *p, __m256 *pChannelWeights)
 {
-    p[0] = _mm256_fmadd_ps(p[0], pWeightsRGB[0], _mm256_fmadd_ps(p[2], pWeightsRGB[1], _mm256_mul_ps(p[4], pWeightsRGB[2])));
-    p[1] = _mm256_fmadd_ps(p[1], pWeightsRGB[0], _mm256_fmadd_ps(p[3], pWeightsRGB[1], _mm256_mul_ps(p[5], pWeightsRGB[2])));
+    p[0] = _mm256_fmadd_ps(p[0], pChannelWeights[0], _mm256_fmadd_ps(p[2], pChannelWeights[1], _mm256_mul_ps(p[4], pChannelWeights[2])));
+    p[1] = _mm256_fmadd_ps(p[1], pChannelWeights[0], _mm256_fmadd_ps(p[3], pChannelWeights[1], _mm256_mul_ps(p[5], pChannelWeights[2])));
 }
 
-inline void compute_rgb_48_to_greyscale_16_host(__m128 *p, __m128 *pWeightsRGB)
+inline void compute_color_48_to_greyscale_16_host(__m128 *p, __m128 *pChannelWeights)
 {
-    p[0] = _mm_fmadd_ps(p[0], pWeightsRGB[0], _mm_fmadd_ps(p[4], pWeightsRGB[1], _mm_mul_ps(p[8], pWeightsRGB[2])));
-    p[1] = _mm_fmadd_ps(p[1], pWeightsRGB[0], _mm_fmadd_ps(p[5], pWeightsRGB[1], _mm_mul_ps(p[9], pWeightsRGB[2])));
-    p[2] = _mm_fmadd_ps(p[2], pWeightsRGB[0], _mm_fmadd_ps(p[6], pWeightsRGB[1], _mm_mul_ps(p[10], pWeightsRGB[2])));
-    p[3] = _mm_fmadd_ps(p[3], pWeightsRGB[0], _mm_fmadd_ps(p[7], pWeightsRGB[1], _mm_mul_ps(p[11], pWeightsRGB[2])));
+    p[0] = _mm_fmadd_ps(p[0], pChannelWeights[0], _mm_fmadd_ps(p[4], pChannelWeights[1], _mm_mul_ps(p[8], pChannelWeights[2])));
+    p[1] = _mm_fmadd_ps(p[1], pChannelWeights[0], _mm_fmadd_ps(p[5], pChannelWeights[1], _mm_mul_ps(p[9], pChannelWeights[2])));
+    p[2] = _mm_fmadd_ps(p[2], pChannelWeights[0], _mm_fmadd_ps(p[6], pChannelWeights[1], _mm_mul_ps(p[10], pChannelWeights[2])));
+    p[3] = _mm_fmadd_ps(p[3], pChannelWeights[0], _mm_fmadd_ps(p[7], pChannelWeights[1], _mm_mul_ps(p[11], pChannelWeights[2])));
 }
 
-inline void compute_rgb_24_to_greyscale_8_host(__m256 *p, __m256 *pWeightsRGB)
+inline void compute_color_24_to_greyscale_8_host(__m256 *p, __m256 *pChannelWeights)
 {
-    p[0] = _mm256_fmadd_ps(p[0], pWeightsRGB[0], _mm256_fmadd_ps(p[1], pWeightsRGB[1], _mm256_mul_ps(p[2], pWeightsRGB[2])));
+    p[0] = _mm256_fmadd_ps(p[0], pChannelWeights[0], _mm256_fmadd_ps(p[1], pChannelWeights[1], _mm256_mul_ps(p[2], pChannelWeights[2])));
 }
 
-inline void compute_rgb_12_to_greyscale_4_host(__m128 *p, __m128 *pWeightsRGB)
+inline void compute_color_12_to_greyscale_4_host(__m128 *p, __m128 *pChannelWeights)
 {
-    p[0] = _mm_fmadd_ps(p[0], pWeightsRGB[0], _mm_fmadd_ps(p[1], pWeightsRGB[1], _mm_mul_ps(p[2], pWeightsRGB[2])));
+    p[0] = _mm_fmadd_ps(p[0], pChannelWeights[0], _mm_fmadd_ps(p[1], pChannelWeights[1], _mm_mul_ps(p[2], pChannelWeights[2])));
 }
 
 inline void compute_contrast_48_host(__m256 *p, __m256 *pContrastParams)

diff --git a/src/modules/cpu/host_tensor_data_exchange_operations.hpp b/src/modules/cpu/host_tensor_data_exchange_operations.hpp
@@ -25,6 +25,6 @@ THE SOFTWARE.
 
 #include "kernel/copy.hpp"
 #include "kernel/swap_channels.hpp"
-#include "kernel/rgb_to_greyscale.hpp"
+#include "kernel/color_to_greyscale.hpp"
 
 #endif // HOST_TENSOR_DATA_EXCHANGEE_OPERATIONS_HPP