aboutsummaryrefslogtreecommitdiffhomepage
path: root/src/opts
diff options
context:
space:
mode:
Diffstat (limited to 'src/opts')
-rw-r--r--src/opts/SkBitmapProcState_opts_SSE2.cpp116
-rw-r--r--src/opts/SkBitmapProcState_opts_SSE2.h3
-rw-r--r--src/opts/SkBlitRow_opts_SSE2.cpp18
-rw-r--r--src/opts/opts_check_SSE2.cpp2
4 files changed, 131 insertions, 8 deletions
diff --git a/src/opts/SkBitmapProcState_opts_SSE2.cpp b/src/opts/SkBitmapProcState_opts_SSE2.cpp
index 7a9e64b1d5..a24c6a159a 100644
--- a/src/opts/SkBitmapProcState_opts_SSE2.cpp
+++ b/src/opts/SkBitmapProcState_opts_SSE2.cpp
@@ -124,3 +124,119 @@ void S32_opaque_D32_filter_DX_SSE2(const SkBitmapProcState& s,
*colors++ = _mm_cvtsi128_si32(sum);
} while (--count > 0);
}
+
+void S32_alpha_D32_filter_DX_SSE2(const SkBitmapProcState& s,
+ const uint32_t* xy,
+ int count, uint32_t* colors) {
+ SkASSERT(count > 0 && colors != NULL);
+ SkASSERT(s.fDoFilter);
+ SkASSERT(s.fBitmap->config() == SkBitmap::kARGB_8888_Config);
+ SkASSERT(s.fAlphaScale < 256);
+
+ const char* srcAddr = static_cast<const char*>(s.fBitmap->getPixels());
+ unsigned rb = s.fBitmap->rowBytes();
+ uint32_t XY = *xy++;
+ unsigned y0 = XY >> 14;
+ const uint32_t* row0 = reinterpret_cast<const uint32_t*>(srcAddr + (y0 >> 4) * rb);
+ const uint32_t* row1 = reinterpret_cast<const uint32_t*>(srcAddr + (XY & 0x3FFF) * rb);
+ unsigned subY = y0 & 0xF;
+
+ // ( 0, 0, 0, 0, 0, 0, 0, 16)
+ __m128i sixteen = _mm_cvtsi32_si128(16);
+
+ // ( 0, 0, 0, 0, 16, 16, 16, 16)
+ sixteen = _mm_shufflelo_epi16(sixteen, 0);
+
+ // ( 0, 0, 0, 0, 0, 0, 0, y)
+ __m128i allY = _mm_cvtsi32_si128(subY);
+
+ // ( 0, 0, 0, 0, y, y, y, y)
+ allY = _mm_shufflelo_epi16(allY, 0);
+
+ // ( 0, 0, 0, 0, 16-y, 16-y, 16-y, 16-y)
+ __m128i negY = _mm_sub_epi16(sixteen, allY);
+
+ // (16-y, 16-y, 16-y, 16-y, y, y, y, y)
+ allY = _mm_unpacklo_epi64(allY, negY);
+
+ // (16, 16, 16, 16, 16, 16, 16, 16 )
+ sixteen = _mm_shuffle_epi32(sixteen, 0);
+
+ // ( 0, 0, 0, 0, 0, 0, 0, 0)
+ __m128i zero = _mm_setzero_si128();
+
+ // ( alpha, alpha, alpha, alpha, alpha, alpha, alpha, alpha )
+ __m128i alpha = _mm_set1_epi16(s.fAlphaScale);
+
+ do {
+ uint32_t XX = *xy++; // x0:14 | 4 | x1:14
+ unsigned x0 = XX >> 18;
+ unsigned x1 = XX & 0x3FFF;
+
+ // (0, 0, 0, 0, 0, 0, 0, x)
+ __m128i allX = _mm_cvtsi32_si128((XX >> 14) & 0x0F);
+
+ // (0, 0, 0, 0, x, x, x, x)
+ allX = _mm_shufflelo_epi16(allX, 0);
+
+ // (x, x, x, x, x, x, x, x)
+ allX = _mm_shuffle_epi32(allX, 0);
+
+ // (16-x, 16-x, 16-x, 16-x, 16-x, 16-x, 16-x)
+ __m128i negX = _mm_sub_epi16(sixteen, allX);
+
+ // Load 4 samples (pixels).
+ __m128i a00 = _mm_cvtsi32_si128(row0[x0]);
+ __m128i a01 = _mm_cvtsi32_si128(row0[x1]);
+ __m128i a10 = _mm_cvtsi32_si128(row1[x0]);
+ __m128i a11 = _mm_cvtsi32_si128(row1[x1]);
+
+ // (0, 0, a00, a10)
+ __m128i a00a10 = _mm_unpacklo_epi32(a10, a00);
+
+ // Expand to 16 bits per component.
+ a00a10 = _mm_unpacklo_epi8(a00a10, zero);
+
+ // ((a00 * (16-y)), (a10 * y)).
+ a00a10 = _mm_mullo_epi16(a00a10, allY);
+
+ // (a00 * (16-y) * (16-x), a10 * y * (16-x)).
+ a00a10 = _mm_mullo_epi16(a00a10, negX);
+
+ // (0, 0, a01, a10)
+ __m128i a01a11 = _mm_unpacklo_epi32(a11, a01);
+
+ // Expand to 16 bits per component.
+ a01a11 = _mm_unpacklo_epi8(a01a11, zero);
+
+ // (a01 * (16-y)), (a11 * y)
+ a01a11 = _mm_mullo_epi16(a01a11, allY);
+
+ // (a01 * (16-y) * x), (a11 * y * x)
+ a01a11 = _mm_mullo_epi16(a01a11, allX);
+
+ // (a00*w00 + a01*w01, a10*w10 + a11*w11)
+ __m128i sum = _mm_add_epi16(a00a10, a01a11);
+
+ // (DC, a00*w00 + a01*w01)
+ __m128i shifted = _mm_shuffle_epi32(sum, 0xEE);
+
+ // (DC, a00*w00 + a01*w01 + a10*w10 + a11*w11)
+ sum = _mm_add_epi16(sum, shifted);
+
+ // Divide each 16 bit component by 256.
+ sum = _mm_srli_epi16(sum, 8);
+
+ // Multiply by alpha.
+ sum = _mm_mullo_epi16(sum, alpha);
+
+ // Divide each 16 bit component by 256.
+ sum = _mm_srli_epi16(sum, 8);
+
+ // Pack lower 4 16 bit values of sum into lower 4 bytes.
+ sum = _mm_packus_epi16(sum, zero);
+
+ // Extract low int and store.
+ *colors++ = _mm_cvtsi128_si32(sum);
+ } while (--count > 0);
+}
diff --git a/src/opts/SkBitmapProcState_opts_SSE2.h b/src/opts/SkBitmapProcState_opts_SSE2.h
index 57342ff7ca..29df88fbd0 100644
--- a/src/opts/SkBitmapProcState_opts_SSE2.h
+++ b/src/opts/SkBitmapProcState_opts_SSE2.h
@@ -20,3 +20,6 @@
void S32_opaque_D32_filter_DX_SSE2(const SkBitmapProcState& s,
const uint32_t* xy,
int count, uint32_t* colors);
+void S32_alpha_D32_filter_DX_SSE2(const SkBitmapProcState& s,
+ const uint32_t* xy,
+ int count, uint32_t* colors);
diff --git a/src/opts/SkBlitRow_opts_SSE2.cpp b/src/opts/SkBlitRow_opts_SSE2.cpp
index bf2db26d79..acf8418b40 100644
--- a/src/opts/SkBlitRow_opts_SSE2.cpp
+++ b/src/opts/SkBlitRow_opts_SSE2.cpp
@@ -125,8 +125,7 @@ void S32A_Opaque_BlitRow32_SSE2(SkPMColor* SK_RESTRICT dst,
__m128i dst_pixel = _mm_load_si128(d);
__m128i dst_rb = _mm_and_si128(rb_mask, dst_pixel);
- __m128i dst_ag = _mm_andnot_si128(rb_mask, dst_pixel);
- dst_ag = _mm_srli_epi16(dst_ag, 8);
+ __m128i dst_ag = _mm_srli_epi16(dst_pixel, 8);
// Shift alphas down to lower 8 bits of each quad.
__m128i alpha = _mm_srli_epi32(src_pixel, 24);
@@ -174,13 +173,16 @@ void S32A_Opaque_BlitRow32_SSE2(SkPMColor* SK_RESTRICT dst,
__m128i dst_pixel = _mm_load_si128(d);
__m128i dst_rb = _mm_and_si128(rb_mask, dst_pixel);
- __m128i dst_ag = _mm_andnot_si128(rb_mask, dst_pixel);
- dst_ag = _mm_srli_epi16(dst_ag, 8);
- // Shift alphas down to lower 8 bits of each quad.
- __m128i alpha = _mm_srli_epi32(src_pixel, 24);
+ __m128i dst_ag = _mm_srli_epi16(dst_pixel, 8);
- // Copy alpha to upper 3rd byte of each quad
- alpha = _mm_or_si128(alpha, _mm_slli_epi32(alpha, 16));
+ // (a0, g0, a1, g1, a2, g2, a3, g3) (low byte of each word)
+ __m128i alpha = _mm_srli_epi16(src_pixel, 8);
+
+ // (a0, a0, a1, a1, a2, g2, a3, g3)
+ alpha = _mm_shufflehi_epi16(alpha, 0xF5);
+
+ // (a0, a0, a1, a1, a2, a2, a3, a3)
+ alpha = _mm_shufflelo_epi16(alpha, 0xF5);
// Subtract alphas from 256, to get 1..256
alpha = _mm_sub_epi16(c_256, alpha);
diff --git a/src/opts/opts_check_SSE2.cpp b/src/opts/opts_check_SSE2.cpp
index 3b84596eaf..6a0112c7df 100644
--- a/src/opts/opts_check_SSE2.cpp
+++ b/src/opts/opts_check_SSE2.cpp
@@ -69,6 +69,8 @@ void SkBitmapProcState::platformProcs() {
if (hasSSE2()) {
if (fSampleProc32 == S32_opaque_D32_filter_DX) {
fSampleProc32 = S32_opaque_D32_filter_DX_SSE2;
+ } else if (fSampleProc32 == S32_alpha_D32_filter_DX) {
+ fSampleProc32 = S32_alpha_D32_filter_DX_SSE2;
}
}
}
generated by cgit v1.2.3 (git 2.39.1) at 2024-12-22 14:59:21 +0000