swr: [rasterizer core/memory] Fix missing avx512 storetile

Fix pre-processor macro handing to eliminate silently missing
implementation for AVX512.

Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>

3 files changed, 14 insertions, 23 deletions

diff --git a/src/gallium/drivers/swr/rasterizer/core/format_types.h b/src/gallium/drivers/swr/rasterizer/core/format_types.h
index faa2e76ded..4c94c312ba 100644
--- a/src/gallium/drivers/swr/rasterizer/core/format_types.h
+++ b/src/gallium/drivers/swr/rasterizer/core/format_types.h
@@ -100,7 +100,7 @@ struct PackTraits<8, false>
     static simdscalar unpack(simdscalar &in)
     {
 #if KNOB_SIMD_WIDTH == 8
-#if KNOB_ARCH==KNOB_ARCH_AVX
+#if KNOB_ARCH <= KNOB_ARCH_AVX
         __m128i src = _mm_castps_si128(_mm256_castps256_ps128(in));
         __m128i resLo = _mm_cvtepu8_epi32(src);
         __m128i resHi = _mm_shuffle_epi8(src,
@@ -109,7 +109,7 @@ struct PackTraits<8, false>
         __m256i result = _mm256_castsi128_si256(resLo);
         result = _mm256_insertf128_si256(result, resHi, 1);
         return _mm256_castsi256_ps(result);
-#elif KNOB_ARCH>=KNOB_ARCH_AVX2
+#else
         return _mm256_castsi256_ps(_mm256_cvtepu8_epi32(_mm_castps_si128(_mm256_castps256_ps128(in))));
 #endif
 #else
@@ -214,7 +214,7 @@ struct PackTraits<8, true>
     static simdscalar unpack(simdscalar &in)
     {
 #if KNOB_SIMD_WIDTH == 8
-#if KNOB_ARCH==KNOB_ARCH_AVX
+#if KNOB_ARCH <= KNOB_ARCH_AVX
         SWR_INVALID("I think this may be incorrect.");
         __m128i src = _mm_castps_si128(_mm256_castps256_ps128(in));
         __m128i resLo = _mm_cvtepi8_epi32(src);
@@ -224,7 +224,7 @@ struct PackTraits<8, true>
         __m256i result = _mm256_castsi128_si256(resLo);
         result = _mm256_insertf128_si256(result, resHi, 1);
         return _mm256_castsi256_ps(result);
-#elif KNOB_ARCH>=KNOB_ARCH_AVX2
+#else
         return _mm256_castsi256_ps(_mm256_cvtepi8_epi32(_mm_castps_si128(_mm256_castps256_ps128(in))));
 #endif
 #else
@@ -329,7 +329,7 @@ struct PackTraits<16, false>
     static simdscalar unpack(simdscalar &in)
     {
 #if KNOB_SIMD_WIDTH == 8
-#if KNOB_ARCH==KNOB_ARCH_AVX
+#if KNOB_ARCH <= KNOB_ARCH_AVX
         __m128i src = _mm_castps_si128(_mm256_castps256_ps128(in));
         __m128i resLo = _mm_cvtepu16_epi32(src);
         __m128i resHi = _mm_shuffle_epi8(src,
@@ -338,7 +338,7 @@ struct PackTraits<16, false>
         __m256i result = _mm256_castsi128_si256(resLo);
         result = _mm256_insertf128_si256(result, resHi, 1);
         return _mm256_castsi256_ps(result);
-#elif KNOB_ARCH>=KNOB_ARCH_AVX2
+#else
         return _mm256_castsi256_ps(_mm256_cvtepu16_epi32(_mm_castps_si128(_mm256_castps256_ps128(in))));
 #endif
 #else
@@ -427,7 +427,7 @@ struct PackTraits<16, true>
     static simdscalar unpack(simdscalar &in)
     {
 #if KNOB_SIMD_WIDTH == 8
-#if KNOB_ARCH==KNOB_ARCH_AVX
+#if KNOB_ARCH <= KNOB_ARCH_AVX
         SWR_INVALID("I think this may be incorrect.");
         __m128i src = _mm_castps_si128(_mm256_castps256_ps128(in));
         __m128i resLo = _mm_cvtepi16_epi32(src);
@@ -437,7 +437,7 @@ struct PackTraits<16, true>
         __m256i result = _mm256_castsi128_si256(resLo);
         result = _mm256_insertf128_si256(result, resHi, 1);
         return _mm256_castsi256_ps(result);
-#elif KNOB_ARCH>=KNOB_ARCH_AVX2
+#else
         return _mm256_castsi256_ps(_mm256_cvtepi16_epi32(_mm_castps_si128(_mm256_castps256_ps128(in))));
 #endif
 #else
@@ -1087,7 +1087,6 @@ template<> struct TypeTraits<SWR_TYPE_FLOAT, 32> : PackTraits<32>
     static inline simdscalar convertSrgb(simdscalar &in)
     {
 #if KNOB_SIMD_WIDTH == 8
-#if (KNOB_ARCH == KNOB_ARCH_AVX || KNOB_ARCH == KNOB_ARCH_AVX2)
         __m128 srcLo = _mm256_extractf128_ps(in, 0);
         __m128 srcHi = _mm256_extractf128_ps(in, 1);
 
@@ -1096,7 +1095,6 @@ template<> struct TypeTraits<SWR_TYPE_FLOAT, 32> : PackTraits<32>
 
         in = _mm256_insertf128_ps(in, srcLo, 0);
         in = _mm256_insertf128_ps(in, srcHi, 1);
-#endif
 #else
 #error Unsupported vector width
 #endif
diff --git a/src/gallium/drivers/swr/rasterizer/core/utils.h b/src/gallium/drivers/swr/rasterizer/core/utils.h
index c4162b4e71..3a0eb25716 100644
--- a/src/gallium/drivers/swr/rasterizer/core/utils.h
+++ b/src/gallium/drivers/swr/rasterizer/core/utils.h
@@ -291,7 +291,7 @@ struct Transpose8_8_8_8
         simdscalari src = _simd_load_si((const simdscalari*)pSrc);
 
 #if KNOB_SIMD_WIDTH == 8
-#if KNOB_ARCH == KNOB_ARCH_AVX
+#if KNOB_ARCH <= KNOB_ARCH_AVX
         __m128i c0c1 = _mm256_castsi256_si128(src);                                           // rrrrrrrrgggggggg
         __m128i c2c3 = _mm_castps_si128(_mm256_extractf128_ps(_mm256_castsi256_ps(src), 1));  // bbbbbbbbaaaaaaaa
         __m128i c0c2 = _mm_unpacklo_epi64(c0c1, c2c3);                                        // rrrrrrrrbbbbbbbb
@@ -302,7 +302,7 @@ struct Transpose8_8_8_8
         __m128i c0123hi = _mm_unpackhi_epi16(c01, c23);                                       // rgbargbargbargba
         _mm_store_si128((__m128i*)pDst, c0123lo);
         _mm_store_si128((__m128i*)(pDst + 16), c0123hi);
-#elif KNOB_ARCH == KNOB_ARCH_AVX2
+#else
         simdscalari dst01 = _mm256_shuffle_epi8(src,
             _mm256_set_epi32(0x0f078080, 0x0e068080, 0x0d058080, 0x0c048080, 0x80800b03, 0x80800a02, 0x80800901, 0x80800800));
         simdscalari dst23 = _mm256_permute2x128_si256(src, src, 0x01);
diff --git a/src/gallium/drivers/swr/rasterizer/memory/StoreTile.h b/src/gallium/drivers/swr/rasterizer/memory/StoreTile.h
index 473ebaef75..f0fdc8c708 100644
--- a/src/gallium/drivers/swr/rasterizer/memory/StoreTile.h
+++ b/src/gallium/drivers/swr/rasterizer/memory/StoreTile.h
@@ -732,7 +732,7 @@ INLINE static void FlatConvert(const uint8_t* pSrc, uint8_t* pDst, uint8_t* pDst
     __m256i src2 = _simd_cvtps_epi32(vComp2); // padded byte bbbbbbbb 
     __m256i src3 = _simd_cvtps_epi32(vComp3); // padded byte aaaaaaaa
 
-#if KNOB_ARCH == KNOB_ARCH_AVX
+#if KNOB_ARCH <= KNOB_ARCH_AVX
 
     // splitting into two sets of 4 wide integer vector types
     // because AVX doesn't have instructions to support this operation at 8 wide
@@ -769,7 +769,7 @@ INLINE static void FlatConvert(const uint8_t* pSrc, uint8_t* pDst, uint8_t* pDst
     __m256i final = _mm256_castsi128_si256(vRow00);
     final = _mm256_insertf128_si256(final, vRow10, 1);
 
-#elif KNOB_ARCH >= KNOB_ARCH_AVX2
+#else
 
     // logic is as above, only wider
     src1 = _mm256_slli_si256(src1, 1);
@@ -780,17 +780,10 @@ INLINE static void FlatConvert(const uint8_t* pSrc, uint8_t* pDst, uint8_t* pDst
     src2 = _mm256_or_si256(src2, src3);
 
     __m256i final = _mm256_or_si256(src0, src2);
-#if 0
-
-    __m256i perm = _mm256_set_epi32(7, 6, 3, 2, 5, 4, 1, 0);
-
-    final = _mm256_permutevar8x32_epi32(final, perm);
-#else
 
     // adjust the data to get the tiling order correct 0 1 2 3 -> 0 2 1 3
     final = _mm256_permute4x64_epi64(final, 0xD8);
 #endif
-#endif
 
     _simd_storeu2_si((__m128i*)pDst1, (__m128i*)pDst, final);
 }
@@ -897,7 +890,7 @@ INLINE static void FlatConvertNoAlpha(const uint8_t* pSrc, uint8_t* pDst, uint8_
     __m256i src1 = _simd_cvtps_epi32(vComp1); // padded byte gggggggg 
     __m256i src2 = _simd_cvtps_epi32(vComp2); // padded byte bbbbbbbb 
 
-#if KNOB_ARCH == KNOB_ARCH_AVX
+#if KNOB_ARCH <= KNOB_ARCH_AVX
 
     // splitting into two sets of 4 wide integer vector types
     // because AVX doesn't have instructions to support this operation at 8 wide
@@ -928,7 +921,7 @@ INLINE static void FlatConvertNoAlpha(const uint8_t* pSrc, uint8_t* pDst, uint8_
     __m256i final = _mm256_castsi128_si256(vRow00);
     final = _mm256_insertf128_si256(final, vRow10, 1);
 
-#elif KNOB_ARCH >= KNOB_ARCH_AVX2
+#else
 
                                               // logic is as above, only wider
     src1 = _mm256_slli_si256(src1, 1);