[Mesa-dev] [PATCH 16/24] swr/rast: SIMD16 FE - fix/use SIMD16 calcDeterminantIntVertical()
Tim Rowley
timothy.o.rowley at intel.com
Sat May 27 21:25:11 UTC 2017
Stop double pumping the SIMD8 version.
---
.../drivers/swr/rasterizer/common/simd16intrin.h | 22 ++++++++
.../drivers/swr/rasterizer/common/simdintrin.h | 24 +++++++++
src/gallium/drivers/swr/rasterizer/core/frontend.h | 62 +++++++---------------
3 files changed, 65 insertions(+), 43 deletions(-)
diff --git a/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h b/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h
index 84585ff..e303ce5 100644
--- a/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h
+++ b/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h
@@ -770,6 +770,26 @@ INLINE simd16scalari SIMDAPI _simd16_cvtepu16_epi32(simdscalari a)
return result;
}
+INLINE simd16scalari SIMDAPI _simd16_cvtepu16_epi64(simdscalari a)
+{
+ simd16scalari result;
+
+ result.lo = _simd_cvtepu16_epi64(_mm256_extractf128_si256(a, 0));
+ result.hi = _simd_cvtepu16_epi64(_mm256_extractf128_si256(a, 1));
+
+ return result;
+}
+
+INLINE simd16scalari SIMDAPI _simd16_cvtepu32_epi64(simdscalari a)
+{
+ simd16scalari result;
+
+ result.lo = _simd_cvtepu32_epi64(_mm256_extractf128_si256(a, 0));
+ result.hi = _simd_cvtepu32_epi64(_mm256_extractf128_si256(a, 1));
+
+ return result;
+}
+
SIMD16_EMU_AVX512_2(simd16scalari, _simd16_packus_epi16, _simd_packus_epi16)
SIMD16_EMU_AVX512_2(simd16scalari, _simd16_packs_epi16, _simd_packs_epi16)
SIMD16_EMU_AVX512_2(simd16scalari, _simd16_packus_epi32, _simd_packus_epi32)
@@ -1097,6 +1117,8 @@ INLINE simd16scalari SIMDAPI _simd16_cmpgt_epi8(simd16scalari a, simd16scalari b
#define _simd16_cvtepu8_epi16 _mm512_cvtepu8_epi16
#define _simd16_cvtepu8_epi32 _mm512_cvtepu8_epi32
#define _simd16_cvtepu16_epi32 _mm512_cvtepu16_epi32
+#define _simd16_cvtepu16_epi64 _mm512_cvtepu16_epi64
+#define _simd16_cvtepu32_epi64 _mm512_cvtepu32_epi64
#define _simd16_packus_epi16 _mm512_packus_epi16
#define _simd16_packs_epi16 _mm512_packs_epi16
#define _simd16_packus_epi32 _mm512_packus_epi32
diff --git a/src/gallium/drivers/swr/rasterizer/common/simdintrin.h b/src/gallium/drivers/swr/rasterizer/common/simdintrin.h
index 61c0c54..ed6e56b 100644
--- a/src/gallium/drivers/swr/rasterizer/common/simdintrin.h
+++ b/src/gallium/drivers/swr/rasterizer/common/simdintrin.h
@@ -456,6 +456,28 @@ __m256i _simd_cvtepu16_epi32(__m128i a)
}
INLINE
+__m256i _simd_cvtepu16_epi64(__m128i a)
+{
+ __m128i resultlo = _mm_cvtepu16_epi64(a);
+ __m128i resulthi = _mm_cvtepu16_epi64(_mm_srli_si128(a, 4));
+
+ __m256i result = _mm256_castsi128_si256(resultlo);
+
+ return _mm256_insertf128_si256(result, resulthi, 1);
+}
+
+INLINE
+__m256i _simd_cvtepu32_epi64(__m128i a)
+{
+ __m128i resultlo = _mm_cvtepu32_epi64(a);
+ __m128i resulthi = _mm_cvtepu32_epi64(_mm_srli_si128(a, 8));
+
+ __m256i result = _mm256_castsi128_si256(resultlo);
+
+ return _mm256_insertf128_si256(result, resulthi, 1);
+}
+
+INLINE
__m256i _simd_packus_epi16(__m256i a, __m256i b)
{
__m128i alo = _mm256_extractf128_si256(a, 0);
@@ -582,6 +604,8 @@ __m256i _simd_packs_epi32(__m256i a, __m256i b)
#define _simd_cvtepu8_epi16 _mm256_cvtepu8_epi16
#define _simd_cvtepu8_epi32 _mm256_cvtepu8_epi32
#define _simd_cvtepu16_epi32 _mm256_cvtepu16_epi32
+#define _simd_cvtepu16_epi64 _mm256_cvtepu16_epi64
+#define _simd_cvtepu32_epi64 _mm256_cvtepu32_epi64
#define _simd_packus_epi16 _mm256_packus_epi16
#define _simd_packs_epi16 _mm256_packs_epi16
#define _simd_packus_epi32 _mm256_packus_epi32
diff --git a/src/gallium/drivers/swr/rasterizer/core/frontend.h b/src/gallium/drivers/swr/rasterizer/core/frontend.h
index eedbcfc..9f347e1 100644
--- a/src/gallium/drivers/swr/rasterizer/core/frontend.h
+++ b/src/gallium/drivers/swr/rasterizer/core/frontend.h
@@ -162,6 +162,7 @@ INLINE
void calcDeterminantIntVertical(const simdscalari vA[3], const simdscalari vB[3], simdscalari *pvDet)
{
// refer to calcDeterminantInt comment for calculation explanation
+
// A1*B2
simdscalari vA1Lo = _simd_unpacklo_epi32(vA[1], vA[1]); // 0 0 1 1 4 4 5 5
simdscalari vA1Hi = _simd_unpackhi_epi32(vA[1], vA[1]); // 2 2 3 3 6 6 7 7
@@ -186,8 +187,10 @@ void calcDeterminantIntVertical(const simdscalari vA[3], const simdscalari vB[3]
simdscalari detLo = _simd_sub_epi64(vA1B2Lo, vA2B1Lo);
simdscalari detHi = _simd_sub_epi64(vA1B2Hi, vA2B1Hi);
- // shuffle 0 1 4 5 -> 0 1 2 3
+ // shuffle 0 1 4 5 2 3 6 7 -> 0 1 2 3
simdscalari vResultLo = _simd_permute2f128_si(detLo, detHi, 0x20);
+
+ // shuffle 0 1 4 5 2 3 6 7 -> 4 5 6 7
simdscalari vResultHi = _simd_permute2f128_si(detLo, detHi, 0x31);
pvDet[0] = vResultLo;
@@ -199,57 +202,30 @@ INLINE
void calcDeterminantIntVertical(const simd16scalari vA[3], const simd16scalari vB[3], simd16scalari *pvDet)
{
// refer to calcDeterminantInt comment for calculation explanation
- // A1*B2
-
-#if 1
- // TODO: get the native SIMD16 version working..
-
- simdscalari vA_lo[3];
- simdscalari vA_hi[3];
- simdscalari vB_lo[3];
- simdscalari vB_hi[3];
-
- for (uint32_t i = 0; i < 3; i += 1)
- {
- vA_lo[i] = _simd16_extract_si(vA[i], 0);
- vA_hi[i] = _simd16_extract_si(vA[i], 1);
- vB_lo[i] = _simd16_extract_si(vB[i], 0);
- vB_hi[i] = _simd16_extract_si(vB[i], 1);
- }
- calcDeterminantIntVertical(vA_lo, vB_lo, reinterpret_cast<simdscalari *>(&pvDet[0]));
- calcDeterminantIntVertical(vA_hi, vB_hi, reinterpret_cast<simdscalari *>(&pvDet[1]));
-#else
- simd16scalari vA1Lo = _simd16_unpacklo_epi32(vA[1], vA[1]); // 0 0 1 1 4 4 5 5 8 8 9 9 C C D D
- simd16scalari vA1Hi = _simd16_unpackhi_epi32(vA[1], vA[1]); // 2 2 3 3 6 6 7 7 A A B B E E F F
+ // A1*B2
+ simd16scalari vA1_lo = _simd16_cvtepu32_epi64(_simd16_extract_si(vA[1], 0));// 0 1 2 3 4 5 6 7 (64b), upper 32b zero, lower 32b used
+ simd16scalari vA1_hi = _simd16_cvtepu32_epi64(_simd16_extract_si(vA[1], 1));// 8 9 A B C D E F (64b), upper 32b zero, lower 32b used
- simd16scalari vB2Lo = _simd16_unpacklo_epi32(vB[2], vB[2]);
- simd16scalari vB2Hi = _simd16_unpackhi_epi32(vB[2], vB[2]);
+ simd16scalari vB2_lo = _simd16_cvtepu32_epi64(_simd16_extract_si(vB[2], 0));// 0 1 2 3 4 5 6 7 (64b), upper 32b zero, lower 32b used
+ simd16scalari vB2_hi = _simd16_cvtepu32_epi64(_simd16_extract_si(vB[2], 1));// 8 9 A B C D E F (64b), upper 32b zero, lower 32b used
- simd16scalari vA1B2Lo = _simd16_mul_epi32(vA1Lo, vB2Lo); // 0 1 4 5 8 9 C D
- simd16scalari vA1B2Hi = _simd16_mul_epi32(vA1Hi, vB2Hi); // 2 3 6 7 A B E F
+ simd16scalari vA1B2_lo = _simd16_mul_epi32(vA1_lo, vB2_lo); // 0 1 2 3 4 5 6 7 (64b)
+ simd16scalari vA1B2_hi = _simd16_mul_epi32(vA1_hi, vB2_hi); // 8 9 A B C D E F (64b)
// B1*A2
- simd16scalari vA2Lo = _simd16_unpacklo_epi32(vA[2], vA[2]);
- simd16scalari vA2Hi = _simd16_unpackhi_epi32(vA[2], vA[2]);
+ simd16scalari vA2_lo = _simd16_cvtepu32_epi64(_simd16_extract_si(vA[2], 0));
+ simd16scalari vA2_hi = _simd16_cvtepu32_epi64(_simd16_extract_si(vA[2], 1));
- simd16scalari vB1Lo = _simd16_unpacklo_epi32(vB[1], vB[1]);
- simd16scalari vB1Hi = _simd16_unpackhi_epi32(vB[1], vB[1]);
+ simd16scalari vB1_lo = _simd16_cvtepu32_epi64(_simd16_extract_si(vB[1], 0));
+ simd16scalari vB1_hi = _simd16_cvtepu32_epi64(_simd16_extract_si(vB[1], 1));
- simd16scalari vA2B1Lo = _simd16_mul_epi32(vA2Lo, vB1Lo);
- simd16scalari vA2B1Hi = _simd16_mul_epi32(vA2Hi, vB1Hi);
+ simd16scalari vA2B1_lo = _simd16_mul_epi32(vA2_lo, vB1_lo);
+ simd16scalari vA2B1_hi = _simd16_mul_epi32(vA2_hi, vB1_hi);
// A1*B2 - A2*B1
- simd16scalari detLo = _simd16_sub_epi64(vA1B2Lo, vA2B1Lo);
- simd16scalari detHi = _simd16_sub_epi64(vA1B2Hi, vA2B1Hi);
-
- // shuffle 0 1 4 5 -> 0 1 2 3
- simd16scalari vResultLo = _simd16_permute2f128_si(detLo, detHi, 0x20);
- simd16scalari vResultHi = _simd16_permute2f128_si(detLo, detHi, 0x31);
-
- pvDet[0] = vResultLo;
- pvDet[1] = vResultHi;
-#endif
+ pvDet[0] = _simd16_sub_epi64(vA1B2_lo, vA2B1_lo); // 0 1 2 3 4 5 6 7 (64b)
+ pvDet[1] = _simd16_sub_epi64(vA1B2_hi, vA2B1_hi); // 8 9 A B C D E F (64b)
}
#endif
--
2.7.4
More information about the mesa-dev
mailing list