swr: [rasterizer core] Enable SIMD16

Make the AVX512 insert/extract intrinsics KNL-compatible

Reviewed-by: George Kyriazis <[email protected]>

3 files changed, 20 insertions, 8 deletions

diff --git a/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h b/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h
index 88814a58aa9..3b43d510e68 100644
--- a/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h
+++ b/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h
@@ -853,10 +853,10 @@ INLINE simd16scalari _simd16_set_epi32(int e7, int e6, int e5, int e4, int e3, i
 #define _simd16_broadcast_ps(m) _mm512_extload_ps(m, _MM_UPCONV_PS_NONE, _MM_BROADCAST_4X16, 0)
 #define _simd16_store_ps        _mm512_store_ps
 #define _simd16_store_si        _mm512_store_si512
-#define _simd16_extract_ps      _mm512_extractf32x8_ps
-#define _simd16_extract_si      _mm512_extracti32x8_epi32
-#define _simd16_insert_ps       _mm512_insertf32x8
-#define _simd16_insert_si       _mm512_inserti32x8
+#define _simd16_extract_ps(a, imm8) _mm256_castsi256_ps(_mm512_extracti64x4_epi64(_mm512_castps_si512(a), imm8))
+#define _simd16_extract_si      _mm512_extracti64x4_epi64
+#define _simd16_insert_ps(a, b, imm8)  _mm512_castsi512_ps(_mm512_inserti64x4(_mm512_castps_si512(a), _mm256_castps_si256(b), imm8))
+#define _simd16_insert_si       _mm512_inserti64x4
 
 INLINE void _simd16_maskstore_ps(float *m, simd16scalari mask, simd16scalar a)
 {
@@ -871,21 +871,21 @@ INLINE simd16scalar _simd16_blendv_ps(simd16scalar a, simd16scalar b, const simd
 {
     simd16mask k = _simd16_scalari2mask(_mm512_castps_si512(mask));
 
-    _mm512_mask_blend_ps(k, a, b);
+    return _mm512_mask_blend_ps(k, a, b);
 }
 
 INLINE simd16scalari _simd16_blendv_epi32(simd16scalari a, simd16scalari b, const simd16scalar mask)
 {
     simd16mask k = _simd16_scalari2mask(_mm512_castps_si512(mask));
 
-    _mm512_mask_blend_epi32(k, a, b);
+    return _mm512_mask_blend_epi32(k, a, b);
 }
 
 INLINE simd16scalari _simd16_blendv_epi32(simd16scalari a, simd16scalari b, const simd16scalari mask)
 {
     simd16mask k = _simd16_scalari2mask(mask);
 
-    _mm512_mask_blend_epi32(k, a, b);
+    return _mm512_mask_blend_epi32(k, a, b);
 }
 
 #define _simd16_mul_ps          _mm512_mul_ps
diff --git a/src/gallium/drivers/swr/rasterizer/core/knobs.h b/src/gallium/drivers/swr/rasterizer/core/knobs.h
index 8e54f90526b..7928f5d6d76 100644
--- a/src/gallium/drivers/swr/rasterizer/core/knobs.h
+++ b/src/gallium/drivers/swr/rasterizer/core/knobs.h
@@ -38,7 +38,7 @@
 // AVX512 Support
 ///////////////////////////////////////////////////////////////////////////////
 
-#define ENABLE_AVX512_SIMD16    0
+#define ENABLE_AVX512_SIMD16    1
 #define USE_8x2_TILE_BACKEND    0
 #define USE_SIMD16_FRONTEND     0
 
diff --git a/src/gallium/drivers/swr/rasterizer/core/pa_avx.cpp b/src/gallium/drivers/swr/rasterizer/core/pa_avx.cpp
index 297f23a88ca..511a1fc0df3 100644
--- a/src/gallium/drivers/swr/rasterizer/core/pa_avx.cpp
+++ b/src/gallium/drivers/swr/rasterizer/core/pa_avx.cpp
@@ -1297,7 +1297,19 @@ bool PaTriFan0_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[])
 
 bool PaTriFan1_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[])
 {
+#if USE_SIMD16_FRONTEND
     const simd16vector &a = pa.leadingVertex.attrib[slot];
+#else
+    simd16vector a;
+
+    {
+        for (uint32_t i = 0; i < 4; i += 1)
+        {
+            a[i] = _simd16_insert_ps(_simd16_setzero_ps(), pa.leadingVertex.attrib[slot][i], 0);
+        }
+    }
+
+#endif
     const simd16vector &b = PaGetSimdVector_simd16(pa, pa.prev, slot);
     const simd16vector &c = PaGetSimdVector_simd16(pa, pa.cur, slot);