Partial fix for padding pixels being counted toward in-range pixels.

Author: Mysticial

SerialPrograms/Source/Kernels/ImageFilters/RGB32_Range/Kernels_ImageFilter_RGB32_Range_x64_AVX2.cpp

@@ -12,6 +12,10 @@
 #include "Kernels/ImageFilters/Kernels_ImageFilter_Basic_Routines.h"
 #include "Kernels_ImageFilter_RGB32_Range.h"
 
+//#include <iostream>
+//using std::cout;
+//using std::endl;
+
 namespace PokemonAutomation{
 namespace Kernels{
 
@@ -42,25 +46,27 @@ class ImageFilterRunner_Rgb32Range_x64_AVX2{
 
     PA_FORCE_INLINE void process_full(uint32_t* out, const uint32_t* in){
         __m256i pixel = _mm256_loadu_si256((const __m256i*)in);
-        pixel = process_word(pixel);
+        __m256i in_range_pixels = process_word(pixel);
+        m_count = _mm256_sub_epi32(m_count, in_range_pixels);
         _mm256_storeu_si256((__m256i*)out, pixel);
     }
     PA_FORCE_INLINE void process_partial(uint32_t* out, const uint32_t* in, const Mask& mask){
         __m256i pixel = mask.load_i32(in);
-        pixel = process_word(pixel);
+        __m256i in_range_pixels = process_word(pixel);
+        in_range_pixels = _mm256_and_si256(in_range_pixels, mask.mask());
+        m_count = _mm256_sub_epi32(m_count, in_range_pixels);
         mask.store(out, pixel);
     }
 
 private:
-    PA_FORCE_INLINE __m256i process_word(__m256i pixel){
+    PA_FORCE_INLINE __m256i process_word(__m256i& pixel){
         __m256i adj = _mm256_xor_si256(pixel, _mm256_set1_epi8((uint8_t)0x80));
         __m256i cmp0 = _mm256_cmpgt_epi8(m_mins, adj);
         __m256i cmp1 = _mm256_cmpgt_epi8(adj, m_maxs);
         cmp0 = _mm256_or_si256(cmp0, cmp1);
         cmp0 = _mm256_cmpeq_epi32(cmp0, _mm256_setzero_si256());
-        m_count = _mm256_sub_epi32(m_count, cmp0);
-        cmp0 = _mm256_xor_si256(cmp0, m_invert);
-        return _mm256_blendv_epi8(m_replacement, pixel, cmp0);
+        pixel = _mm256_blendv_epi8(m_replacement, pixel, _mm256_xor_si256(cmp0, m_invert));
+        return cmp0;
     }
 
 private:

SerialPrograms/Source/Kernels/ImageFilters/RGB32_Range/Kernels_ImageFilter_RGB32_Range_x64_AVX512.cpp

@@ -10,6 +10,10 @@
 #include "Kernels/ImageFilters/Kernels_ImageFilter_Basic_Routines.h"
 #include "Kernels_ImageFilter_RGB32_Range.h"
 
+//#include <iostream>
+//using std::cout;
+//using std::endl;
+
 namespace PokemonAutomation{
 namespace Kernels{
 
@@ -55,17 +59,20 @@ class ImageFilterRunner_Rgb32Range_x64_AVX512{
 
     PA_FORCE_INLINE void process_full(uint32_t* out, const uint32_t* in){
         __m512i pixel = _mm512_loadu_si512((const __m512i*)in);
-        pixel = process_word(pixel);
+        __mmask16 in_range_pixels = process_word(pixel);
+        m_count = _mm512_mask_sub_epi32(m_count, in_range_pixels, m_count, _mm512_set1_epi32(-1));
         _mm512_storeu_si512((__m512i*)out, pixel);
     }
     PA_FORCE_INLINE void process_partial(uint32_t* out, const uint32_t* in, const Mask& mask){
         __m512i pixel = _mm512_maskz_loadu_epi32(mask.m, in);
-        pixel = process_word(pixel);
+        __mmask16 in_range_pixels = process_word(pixel);
+        in_range_pixels &= mask.m;
+        m_count = _mm512_mask_sub_epi32(m_count, in_range_pixels, m_count, _mm512_set1_epi32(-1));
         _mm512_mask_storeu_epi32(out, mask.m, pixel);
     }
 
 private:
-    PA_FORCE_INLINE __m512i process_word(__m512i pixel){
+    PA_FORCE_INLINE __mmask16 process_word(__m512i& pixel){
 #if 0
         __mmask64 cmp64A = _mm512_cmpgt_epu8_mask(m_mins, pixel);
         __mmask64 cmp64B = _mm512_cmpgt_epu8_mask(pixel, m_maxs);
@@ -77,9 +84,9 @@ class ImageFilterRunner_Rgb32Range_x64_AVX512{
         __m512i mask = _mm512_movm_epi8(cmp64B);
         __mmask16 cmp16 = _mm512_cmpeq_epi32_mask(mask, _mm512_set1_epi32(-1));
 #endif
-        m_count = _mm512_mask_sub_epi32(m_count, cmp16, m_count, _mm512_set1_epi32(-1));
-        cmp16 ^= m_invert;
-        return _mm512_mask_blend_epi32(cmp16, m_replacement, pixel);
+        pixel = _mm512_mask_blend_epi32(cmp16 ^ m_invert, m_replacement, pixel);
+
+        return cmp16;
     }
 
 private:

SerialPrograms/Source/Kernels/ImageFilters/RGB32_Range/Kernels_ImageFilter_RGB32_Range_x64_SSE42.cpp

@@ -12,6 +12,10 @@
 #include "Kernels/ImageFilters/Kernels_ImageFilter_Basic_Routines.h"
 #include "Kernels_ImageFilter_RGB32_Range.h"
 
+//#include <iostream>
+//using std::cout;
+//using std::endl;
+
 namespace PokemonAutomation{
 namespace Kernels{
 
@@ -59,12 +63,20 @@ class ImageFilterRunner_Rgb32Range_x64_SSE42{
 
     PA_FORCE_INLINE void process_full(uint32_t* out, const uint32_t* in){
         __m128i pixel = _mm_loadu_si128((const __m128i*)in);
-        pixel = process_word(pixel);
+        __m128i in_range_pixels = process_word(pixel);
+        m_count = _mm_sub_epi32(m_count, in_range_pixels);
         _mm_storeu_si128((__m128i*)out, pixel);
     }
     PA_FORCE_INLINE void process_partial(uint32_t* out, const uint32_t* in, const Mask& mask){
+        __m128i vmask = _mm_cmpgt_epi32(
+            _mm_set1_epi32((uint32_t)mask.left),
+            _mm_setr_epi32(0, 1, 2, 3)
+        );
+
         __m128i pixel = mask.loader.load(in);
-        pixel = process_word(pixel);
+        __m128i in_range_pixels = process_word(pixel);
+        in_range_pixels = _mm_and_si128(in_range_pixels, vmask);
+        m_count = _mm_sub_epi32(m_count, in_range_pixels);
         size_t left = mask.left;
         do{
             out[0] = _mm_cvtsi128_si32(pixel);
@@ -74,15 +86,14 @@ class ImageFilterRunner_Rgb32Range_x64_SSE42{
     }
 
 private:
-    PA_FORCE_INLINE __m128i process_word(__m128i pixel){
+    PA_FORCE_INLINE __m128i process_word(__m128i& pixel){
         __m128i adj = _mm_xor_si128(pixel, _mm_set1_epi8((uint8_t)0x80));
         __m128i cmp0 = _mm_cmpgt_epi8(m_mins, adj);
         __m128i cmp1 = _mm_cmpgt_epi8(adj, m_maxs);
         cmp0 = _mm_or_si128(cmp0, cmp1);
         cmp0 = _mm_cmpeq_epi32(cmp0, _mm_setzero_si128());
-        m_count = _mm_sub_epi32(m_count, cmp0);
-        cmp0 = _mm_xor_si128(cmp0, m_invert);
-        return _mm_blendv_epi8(m_replacement, pixel, cmp0);
+        pixel = _mm_blendv_epi8(m_replacement, pixel, _mm_xor_si128(cmp0, m_invert));
+        return cmp0;
     }
 
 private:

SerialPrograms/Source/Kernels/PartialWordAccess/Kernels_PartialWordAccess_x64_AVX2.h

@@ -25,6 +25,9 @@ class PartialWordAccess32_x64_AVX2{
         );
     }
 
+    PA_FORCE_INLINE __m256i mask() const{
+        return m_mask;
+    }
     PA_FORCE_INLINE __m256i load_i32(const void* ptr) const{
         return _mm256_maskload_epi32((const int*)ptr, m_mask);
     }