summaryrefslogtreecommitdiff
path: root/drivers/webp/dsp/yuv_sse2.c
diff options
context:
space:
mode:
Diffstat (limited to 'drivers/webp/dsp/yuv_sse2.c')
-rw-r--r--drivers/webp/dsp/yuv_sse2.c677
1 files changed, 423 insertions, 254 deletions
diff --git a/drivers/webp/dsp/yuv_sse2.c b/drivers/webp/dsp/yuv_sse2.c
index 283b3af228..e19bddff6c 100644
--- a/drivers/webp/dsp/yuv_sse2.c
+++ b/drivers/webp/dsp/yuv_sse2.c
@@ -16,172 +16,294 @@
#if defined(WEBP_USE_SSE2)
#include <emmintrin.h>
-#include <string.h> // for memcpy
-typedef union { // handy struct for converting SSE2 registers
- int32_t i32[4];
- uint8_t u8[16];
- __m128i m;
-} VP8kCstSSE2;
-
-#if defined(WEBP_YUV_USE_SSE2_TABLES)
-
-#include "./yuv_tables_sse2.h"
-
-WEBP_TSAN_IGNORE_FUNCTION void VP8YUVInitSSE2(void) {}
+//-----------------------------------------------------------------------------
+// Convert spans of 32 pixels to various RGB formats for the fancy upsampler.
-#else
+// These constants are 14b fixed-point version of ITU-R BT.601 constants.
+// R = (19077 * y + 26149 * v - 14234) >> 6
+// G = (19077 * y - 6419 * u - 13320 * v + 8708) >> 6
+// B = (19077 * y + 33050 * u - 17685) >> 6
+static void ConvertYUV444ToRGB(const __m128i* const Y0,
+ const __m128i* const U0,
+ const __m128i* const V0,
+ __m128i* const R,
+ __m128i* const G,
+ __m128i* const B) {
+ const __m128i k19077 = _mm_set1_epi16(19077);
+ const __m128i k26149 = _mm_set1_epi16(26149);
+ const __m128i k14234 = _mm_set1_epi16(14234);
+ // 33050 doesn't fit in a signed short: only use this with unsigned arithmetic
+ const __m128i k33050 = _mm_set1_epi16((short)33050);
+ const __m128i k17685 = _mm_set1_epi16(17685);
+ const __m128i k6419 = _mm_set1_epi16(6419);
+ const __m128i k13320 = _mm_set1_epi16(13320);
+ const __m128i k8708 = _mm_set1_epi16(8708);
+
+ const __m128i Y1 = _mm_mulhi_epu16(*Y0, k19077);
+
+ const __m128i R0 = _mm_mulhi_epu16(*V0, k26149);
+ const __m128i R1 = _mm_sub_epi16(Y1, k14234);
+ const __m128i R2 = _mm_add_epi16(R1, R0);
+
+ const __m128i G0 = _mm_mulhi_epu16(*U0, k6419);
+ const __m128i G1 = _mm_mulhi_epu16(*V0, k13320);
+ const __m128i G2 = _mm_add_epi16(Y1, k8708);
+ const __m128i G3 = _mm_add_epi16(G0, G1);
+ const __m128i G4 = _mm_sub_epi16(G2, G3);
+
+ // be careful with the saturated *unsigned* arithmetic here!
+ const __m128i B0 = _mm_mulhi_epu16(*U0, k33050);
+ const __m128i B1 = _mm_adds_epu16(B0, Y1);
+ const __m128i B2 = _mm_subs_epu16(B1, k17685);
+
+ // use logical shift for B2, which can be larger than 32767
+ *R = _mm_srai_epi16(R2, 6); // range: [-14234, 30815]
+ *G = _mm_srai_epi16(G4, 6); // range: [-10953, 27710]
+ *B = _mm_srli_epi16(B2, 6); // range: [0, 34238]
+}
-static int done_sse2 = 0;
-static VP8kCstSSE2 VP8kUtoRGBA[256], VP8kVtoRGBA[256], VP8kYtoRGBA[256];
-
-WEBP_TSAN_IGNORE_FUNCTION void VP8YUVInitSSE2(void) {
- if (!done_sse2) {
- int i;
- for (i = 0; i < 256; ++i) {
- VP8kYtoRGBA[i].i32[0] =
- VP8kYtoRGBA[i].i32[1] =
- VP8kYtoRGBA[i].i32[2] = (i - 16) * kYScale + YUV_HALF2;
- VP8kYtoRGBA[i].i32[3] = 0xff << YUV_FIX2;
-
- VP8kUtoRGBA[i].i32[0] = 0;
- VP8kUtoRGBA[i].i32[1] = -kUToG * (i - 128);
- VP8kUtoRGBA[i].i32[2] = kUToB * (i - 128);
- VP8kUtoRGBA[i].i32[3] = 0;
-
- VP8kVtoRGBA[i].i32[0] = kVToR * (i - 128);
- VP8kVtoRGBA[i].i32[1] = -kVToG * (i - 128);
- VP8kVtoRGBA[i].i32[2] = 0;
- VP8kVtoRGBA[i].i32[3] = 0;
- }
- done_sse2 = 1;
-
-#if 0 // code used to generate 'yuv_tables_sse2.h'
- printf("static const VP8kCstSSE2 VP8kYtoRGBA[256] = {\n");
- for (i = 0; i < 256; ++i) {
- printf(" {{0x%.8x, 0x%.8x, 0x%.8x, 0x%.8x}},\n",
- VP8kYtoRGBA[i].i32[0], VP8kYtoRGBA[i].i32[1],
- VP8kYtoRGBA[i].i32[2], VP8kYtoRGBA[i].i32[3]);
- }
- printf("};\n\n");
- printf("static const VP8kCstSSE2 VP8kUtoRGBA[256] = {\n");
- for (i = 0; i < 256; ++i) {
- printf(" {{0, 0x%.8x, 0x%.8x, 0}},\n",
- VP8kUtoRGBA[i].i32[1], VP8kUtoRGBA[i].i32[2]);
- }
- printf("};\n\n");
- printf("static VP8kCstSSE2 VP8kVtoRGBA[256] = {\n");
- for (i = 0; i < 256; ++i) {
- printf(" {{0x%.8x, 0x%.8x, 0, 0}},\n",
- VP8kVtoRGBA[i].i32[0], VP8kVtoRGBA[i].i32[1]);
- }
- printf("};\n\n");
-#endif
- }
+// Load the bytes into the *upper* part of 16b words. That's "<< 8", basically.
+static WEBP_INLINE __m128i Load_HI_16(const uint8_t* src) {
+ const __m128i zero = _mm_setzero_si128();
+ return _mm_unpacklo_epi8(zero, _mm_loadl_epi64((const __m128i*)src));
}
-#endif // WEBP_YUV_USE_SSE2_TABLES
+// Load and replicate the U/V samples
+static WEBP_INLINE __m128i Load_UV_HI_8(const uint8_t* src) {
+ const __m128i zero = _mm_setzero_si128();
+ const __m128i tmp0 = _mm_cvtsi32_si128(*(const uint32_t*)src);
+ const __m128i tmp1 = _mm_unpacklo_epi8(zero, tmp0);
+ return _mm_unpacklo_epi16(tmp1, tmp1); // replicate samples
+}
-//-----------------------------------------------------------------------------
+// Convert 32 samples of YUV444 to R/G/B
+static void YUV444ToRGB(const uint8_t* const y,
+ const uint8_t* const u,
+ const uint8_t* const v,
+ __m128i* const R, __m128i* const G, __m128i* const B) {
+ const __m128i Y0 = Load_HI_16(y), U0 = Load_HI_16(u), V0 = Load_HI_16(v);
+ ConvertYUV444ToRGB(&Y0, &U0, &V0, R, G, B);
+}
-static WEBP_INLINE __m128i LoadUVPart(int u, int v) {
- const __m128i u_part = _mm_loadu_si128(&VP8kUtoRGBA[u].m);
- const __m128i v_part = _mm_loadu_si128(&VP8kVtoRGBA[v].m);
- const __m128i uv_part = _mm_add_epi32(u_part, v_part);
- return uv_part;
+// Convert 32 samples of YUV420 to R/G/B
+static void YUV420ToRGB(const uint8_t* const y,
+ const uint8_t* const u,
+ const uint8_t* const v,
+ __m128i* const R, __m128i* const G, __m128i* const B) {
+ const __m128i Y0 = Load_HI_16(y), U0 = Load_UV_HI_8(u), V0 = Load_UV_HI_8(v);
+ ConvertYUV444ToRGB(&Y0, &U0, &V0, R, G, B);
}
-static WEBP_INLINE __m128i GetRGBA32bWithUV(int y, const __m128i uv_part) {
- const __m128i y_part = _mm_loadu_si128(&VP8kYtoRGBA[y].m);
- const __m128i rgba1 = _mm_add_epi32(y_part, uv_part);
- const __m128i rgba2 = _mm_srai_epi32(rgba1, YUV_FIX2);
- return rgba2;
+// Pack R/G/B/A results into 32b output.
+static WEBP_INLINE void PackAndStore4(const __m128i* const R,
+ const __m128i* const G,
+ const __m128i* const B,
+ const __m128i* const A,
+ uint8_t* const dst) {
+ const __m128i rb = _mm_packus_epi16(*R, *B);
+ const __m128i ga = _mm_packus_epi16(*G, *A);
+ const __m128i rg = _mm_unpacklo_epi8(rb, ga);
+ const __m128i ba = _mm_unpackhi_epi8(rb, ga);
+ const __m128i RGBA_lo = _mm_unpacklo_epi16(rg, ba);
+ const __m128i RGBA_hi = _mm_unpackhi_epi16(rg, ba);
+ _mm_storeu_si128((__m128i*)(dst + 0), RGBA_lo);
+ _mm_storeu_si128((__m128i*)(dst + 16), RGBA_hi);
}
-static WEBP_INLINE __m128i GetRGBA32b(int y, int u, int v) {
- const __m128i uv_part = LoadUVPart(u, v);
- return GetRGBA32bWithUV(y, uv_part);
+// Pack R/G/B/A results into 16b output.
+static WEBP_INLINE void PackAndStore4444(const __m128i* const R,
+ const __m128i* const G,
+ const __m128i* const B,
+ const __m128i* const A,
+ uint8_t* const dst) {
+#if !defined(WEBP_SWAP_16BIT_CSP)
+ const __m128i rg0 = _mm_packus_epi16(*R, *G);
+ const __m128i ba0 = _mm_packus_epi16(*B, *A);
+#else
+ const __m128i rg0 = _mm_packus_epi16(*B, *A);
+ const __m128i ba0 = _mm_packus_epi16(*R, *G);
+#endif
+ const __m128i mask_0xf0 = _mm_set1_epi8(0xf0);
+ const __m128i rb1 = _mm_unpacklo_epi8(rg0, ba0); // rbrbrbrbrb...
+ const __m128i ga1 = _mm_unpackhi_epi8(rg0, ba0); // gagagagaga...
+ const __m128i rb2 = _mm_and_si128(rb1, mask_0xf0);
+ const __m128i ga2 = _mm_srli_epi16(_mm_and_si128(ga1, mask_0xf0), 4);
+ const __m128i rgba4444 = _mm_or_si128(rb2, ga2);
+ _mm_storeu_si128((__m128i*)dst, rgba4444);
}
-static WEBP_INLINE void YuvToRgbSSE2(uint8_t y, uint8_t u, uint8_t v,
- uint8_t* const rgb) {
- const __m128i tmp0 = GetRGBA32b(y, u, v);
- const __m128i tmp1 = _mm_packs_epi32(tmp0, tmp0);
- const __m128i tmp2 = _mm_packus_epi16(tmp1, tmp1);
- // Note: we store 8 bytes at a time, not 3 bytes! -> memory stomp
- _mm_storel_epi64((__m128i*)rgb, tmp2);
+// Pack R/G/B results into 16b output.
+static WEBP_INLINE void PackAndStore565(const __m128i* const R,
+ const __m128i* const G,
+ const __m128i* const B,
+ uint8_t* const dst) {
+ const __m128i r0 = _mm_packus_epi16(*R, *R);
+ const __m128i g0 = _mm_packus_epi16(*G, *G);
+ const __m128i b0 = _mm_packus_epi16(*B, *B);
+ const __m128i r1 = _mm_and_si128(r0, _mm_set1_epi8(0xf8));
+ const __m128i b1 = _mm_and_si128(_mm_srli_epi16(b0, 3), _mm_set1_epi8(0x1f));
+ const __m128i g1 = _mm_srli_epi16(_mm_and_si128(g0, _mm_set1_epi8(0xe0)), 5);
+ const __m128i g2 = _mm_slli_epi16(_mm_and_si128(g0, _mm_set1_epi8(0x1c)), 3);
+ const __m128i rg = _mm_or_si128(r1, g1);
+ const __m128i gb = _mm_or_si128(g2, b1);
+#if !defined(WEBP_SWAP_16BIT_CSP)
+ const __m128i rgb565 = _mm_unpacklo_epi8(rg, gb);
+#else
+ const __m128i rgb565 = _mm_unpacklo_epi8(gb, rg);
+#endif
+ _mm_storeu_si128((__m128i*)dst, rgb565);
}
-static WEBP_INLINE void YuvToBgrSSE2(uint8_t y, uint8_t u, uint8_t v,
- uint8_t* const bgr) {
- const __m128i tmp0 = GetRGBA32b(y, u, v);
- const __m128i tmp1 = _mm_shuffle_epi32(tmp0, _MM_SHUFFLE(3, 0, 1, 2));
- const __m128i tmp2 = _mm_packs_epi32(tmp1, tmp1);
- const __m128i tmp3 = _mm_packus_epi16(tmp2, tmp2);
- // Note: we store 8 bytes at a time, not 3 bytes! -> memory stomp
- _mm_storel_epi64((__m128i*)bgr, tmp3);
+// Function used several times in PlanarTo24b.
+// It samples the in buffer as follows: one every two unsigned char is stored
+// at the beginning of the buffer, while the other half is stored at the end.
+static WEBP_INLINE void PlanarTo24bHelper(const __m128i* const in /*in[6]*/,
+ __m128i* const out /*out[6]*/) {
+ const __m128i v_mask = _mm_set1_epi16(0x00ff);
+
+ // Take one every two upper 8b values.
+ out[0] = _mm_packus_epi16(_mm_and_si128(in[0], v_mask),
+ _mm_and_si128(in[1], v_mask));
+ out[1] = _mm_packus_epi16(_mm_and_si128(in[2], v_mask),
+ _mm_and_si128(in[3], v_mask));
+ out[2] = _mm_packus_epi16(_mm_and_si128(in[4], v_mask),
+ _mm_and_si128(in[5], v_mask));
+ // Take one every two lower 8b values.
+ out[3] = _mm_packus_epi16(_mm_srli_epi16(in[0], 8), _mm_srli_epi16(in[1], 8));
+ out[4] = _mm_packus_epi16(_mm_srli_epi16(in[2], 8), _mm_srli_epi16(in[3], 8));
+ out[5] = _mm_packus_epi16(_mm_srli_epi16(in[4], 8), _mm_srli_epi16(in[5], 8));
}
-//-----------------------------------------------------------------------------
-// Convert spans of 32 pixels to various RGB formats for the fancy upsampler.
+// Pack the planar buffers
+// rrrr... rrrr... gggg... gggg... bbbb... bbbb....
+// triplet by triplet in the output buffer rgb as rgbrgbrgbrgb ...
+static WEBP_INLINE void PlanarTo24b(__m128i* const in /*in[6]*/, uint8_t* rgb) {
+ // The input is 6 registers of sixteen 8b but for the sake of explanation,
+ // let's take 6 registers of four 8b values.
+ // To pack, we will keep taking one every two 8b integer and move it
+ // around as follows:
+ // Input:
+ // r0r1r2r3 | r4r5r6r7 | g0g1g2g3 | g4g5g6g7 | b0b1b2b3 | b4b5b6b7
+ // Split the 6 registers in two sets of 3 registers: the first set as the even
+ // 8b bytes, the second the odd ones:
+ // r0r2r4r6 | g0g2g4g6 | b0b2b4b6 | r1r3r5r7 | g1g3g5g7 | b1b3b5b7
+ // Repeat the same permutations twice more:
+ // r0r4g0g4 | b0b4r1r5 | g1g5b1b5 | r2r6g2g6 | b2b6r3r7 | g3g7b3b7
+ // r0g0b0r1 | g1b1r2g2 | b2r3g3b3 | r4g4b4r5 | g5b5r6g6 | b6r7g7b7
+ __m128i tmp[6];
+ PlanarTo24bHelper(in, tmp);
+ PlanarTo24bHelper(tmp, in);
+ PlanarTo24bHelper(in, tmp);
+ // We need to do it two more times than the example as we have sixteen bytes.
+ PlanarTo24bHelper(tmp, in);
+ PlanarTo24bHelper(in, tmp);
+
+ _mm_storeu_si128((__m128i*)(rgb + 0), tmp[0]);
+ _mm_storeu_si128((__m128i*)(rgb + 16), tmp[1]);
+ _mm_storeu_si128((__m128i*)(rgb + 32), tmp[2]);
+ _mm_storeu_si128((__m128i*)(rgb + 48), tmp[3]);
+ _mm_storeu_si128((__m128i*)(rgb + 64), tmp[4]);
+ _mm_storeu_si128((__m128i*)(rgb + 80), tmp[5]);
+}
+#undef MK_UINT32
void VP8YuvToRgba32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
uint8_t* dst) {
+ const __m128i kAlpha = _mm_set1_epi16(255);
int n;
- for (n = 0; n < 32; n += 4) {
- const __m128i tmp0_1 = GetRGBA32b(y[n + 0], u[n + 0], v[n + 0]);
- const __m128i tmp0_2 = GetRGBA32b(y[n + 1], u[n + 1], v[n + 1]);
- const __m128i tmp0_3 = GetRGBA32b(y[n + 2], u[n + 2], v[n + 2]);
- const __m128i tmp0_4 = GetRGBA32b(y[n + 3], u[n + 3], v[n + 3]);
- const __m128i tmp1_1 = _mm_packs_epi32(tmp0_1, tmp0_2);
- const __m128i tmp1_2 = _mm_packs_epi32(tmp0_3, tmp0_4);
- const __m128i tmp2 = _mm_packus_epi16(tmp1_1, tmp1_2);
- _mm_storeu_si128((__m128i*)dst, tmp2);
- dst += 4 * 4;
+ for (n = 0; n < 32; n += 8, dst += 32) {
+ __m128i R, G, B;
+ YUV444ToRGB(y + n, u + n, v + n, &R, &G, &B);
+ PackAndStore4(&R, &G, &B, &kAlpha, dst);
}
}
void VP8YuvToBgra32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
uint8_t* dst) {
+ const __m128i kAlpha = _mm_set1_epi16(255);
int n;
- for (n = 0; n < 32; n += 2) {
- const __m128i tmp0_1 = GetRGBA32b(y[n + 0], u[n + 0], v[n + 0]);
- const __m128i tmp0_2 = GetRGBA32b(y[n + 1], u[n + 1], v[n + 1]);
- const __m128i tmp1_1 = _mm_shuffle_epi32(tmp0_1, _MM_SHUFFLE(3, 0, 1, 2));
- const __m128i tmp1_2 = _mm_shuffle_epi32(tmp0_2, _MM_SHUFFLE(3, 0, 1, 2));
- const __m128i tmp2_1 = _mm_packs_epi32(tmp1_1, tmp1_2);
- const __m128i tmp3 = _mm_packus_epi16(tmp2_1, tmp2_1);
- _mm_storel_epi64((__m128i*)dst, tmp3);
- dst += 4 * 2;
+ for (n = 0; n < 32; n += 8, dst += 32) {
+ __m128i R, G, B;
+ YUV444ToRGB(y + n, u + n, v + n, &R, &G, &B);
+ PackAndStore4(&B, &G, &R, &kAlpha, dst);
}
}
-void VP8YuvToRgb32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
- uint8_t* dst) {
+void VP8YuvToArgb32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+ uint8_t* dst) {
+ const __m128i kAlpha = _mm_set1_epi16(255);
int n;
- uint8_t tmp0[2 * 3 + 5 + 15];
- uint8_t* const tmp = (uint8_t*)((uintptr_t)(tmp0 + 15) & ~15); // align
- for (n = 0; n < 30; ++n) { // we directly stomp the *dst memory
- YuvToRgbSSE2(y[n], u[n], v[n], dst + n * 3);
+ for (n = 0; n < 32; n += 8, dst += 32) {
+ __m128i R, G, B;
+ YUV444ToRGB(y + n, u + n, v + n, &R, &G, &B);
+ PackAndStore4(&kAlpha, &R, &G, &B, dst);
}
- // Last two pixels are special: we write in a tmp buffer before sending
- // to dst.
- YuvToRgbSSE2(y[n + 0], u[n + 0], v[n + 0], tmp + 0);
- YuvToRgbSSE2(y[n + 1], u[n + 1], v[n + 1], tmp + 3);
- memcpy(dst + n * 3, tmp, 2 * 3);
}
-void VP8YuvToBgr32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
- uint8_t* dst) {
+void VP8YuvToRgba444432(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+ uint8_t* dst) {
+ const __m128i kAlpha = _mm_set1_epi16(255);
int n;
- uint8_t tmp0[2 * 3 + 5 + 15];
- uint8_t* const tmp = (uint8_t*)((uintptr_t)(tmp0 + 15) & ~15); // align
- for (n = 0; n < 30; ++n) {
- YuvToBgrSSE2(y[n], u[n], v[n], dst + n * 3);
+ for (n = 0; n < 32; n += 8, dst += 16) {
+ __m128i R, G, B;
+ YUV444ToRGB(y + n, u + n, v + n, &R, &G, &B);
+ PackAndStore4444(&R, &G, &B, &kAlpha, dst);
}
- YuvToBgrSSE2(y[n + 0], u[n + 0], v[n + 0], tmp + 0);
- YuvToBgrSSE2(y[n + 1], u[n + 1], v[n + 1], tmp + 3);
- memcpy(dst + n * 3, tmp, 2 * 3);
+}
+
+void VP8YuvToRgb56532(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+ uint8_t* dst) {
+ int n;
+ for (n = 0; n < 32; n += 8, dst += 16) {
+ __m128i R, G, B;
+ YUV444ToRGB(y + n, u + n, v + n, &R, &G, &B);
+ PackAndStore565(&R, &G, &B, dst);
+ }
+}
+
+void VP8YuvToRgb32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+ uint8_t* dst) {
+ __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
+ __m128i rgb[6];
+
+ YUV444ToRGB(y + 0, u + 0, v + 0, &R0, &G0, &B0);
+ YUV444ToRGB(y + 8, u + 8, v + 8, &R1, &G1, &B1);
+ YUV444ToRGB(y + 16, u + 16, v + 16, &R2, &G2, &B2);
+ YUV444ToRGB(y + 24, u + 24, v + 24, &R3, &G3, &B3);
+
+ // Cast to 8b and store as RRRRGGGGBBBB.
+ rgb[0] = _mm_packus_epi16(R0, R1);
+ rgb[1] = _mm_packus_epi16(R2, R3);
+ rgb[2] = _mm_packus_epi16(G0, G1);
+ rgb[3] = _mm_packus_epi16(G2, G3);
+ rgb[4] = _mm_packus_epi16(B0, B1);
+ rgb[5] = _mm_packus_epi16(B2, B3);
+
+ // Pack as RGBRGBRGBRGB.
+ PlanarTo24b(rgb, dst);
+}
+
+void VP8YuvToBgr32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+ uint8_t* dst) {
+ __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
+ __m128i bgr[6];
+
+ YUV444ToRGB(y + 0, u + 0, v + 0, &R0, &G0, &B0);
+ YUV444ToRGB(y + 8, u + 8, v + 8, &R1, &G1, &B1);
+ YUV444ToRGB(y + 16, u + 16, v + 16, &R2, &G2, &B2);
+ YUV444ToRGB(y + 24, u + 24, v + 24, &R3, &G3, &B3);
+
+ // Cast to 8b and store as BBBBGGGGRRRR.
+ bgr[0] = _mm_packus_epi16(B0, B1);
+ bgr[1] = _mm_packus_epi16(B2, B3);
+ bgr[2] = _mm_packus_epi16(G0, G1);
+ bgr[3] = _mm_packus_epi16(G2, G3);
+ bgr[4] = _mm_packus_epi16(R0, R1);
+ bgr[5] = _mm_packus_epi16(R2, R3);
+
+ // Pack as BGRBGRBGRBGR.
+ PlanarTo24b(bgr, dst);
}
//-----------------------------------------------------------------------------
@@ -189,110 +311,137 @@ void VP8YuvToBgr32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
static void YuvToRgbaRow(const uint8_t* y, const uint8_t* u, const uint8_t* v,
uint8_t* dst, int len) {
+ const __m128i kAlpha = _mm_set1_epi16(255);
int n;
- for (n = 0; n + 4 <= len; n += 4) {
- const __m128i uv_0 = LoadUVPart(u[0], v[0]);
- const __m128i uv_1 = LoadUVPart(u[1], v[1]);
- const __m128i tmp0_1 = GetRGBA32bWithUV(y[0], uv_0);
- const __m128i tmp0_2 = GetRGBA32bWithUV(y[1], uv_0);
- const __m128i tmp0_3 = GetRGBA32bWithUV(y[2], uv_1);
- const __m128i tmp0_4 = GetRGBA32bWithUV(y[3], uv_1);
- const __m128i tmp1_1 = _mm_packs_epi32(tmp0_1, tmp0_2);
- const __m128i tmp1_2 = _mm_packs_epi32(tmp0_3, tmp0_4);
- const __m128i tmp2 = _mm_packus_epi16(tmp1_1, tmp1_2);
- _mm_storeu_si128((__m128i*)dst, tmp2);
- dst += 4 * 4;
- y += 4;
- u += 2;
- v += 2;
+ for (n = 0; n + 8 <= len; n += 8, dst += 32) {
+ __m128i R, G, B;
+ YUV420ToRGB(y, u, v, &R, &G, &B);
+ PackAndStore4(&R, &G, &B, &kAlpha, dst);
+ y += 8;
+ u += 4;
+ v += 4;
}
- // Finish off
- while (n < len) {
+ for (; n < len; ++n) { // Finish off
VP8YuvToRgba(y[0], u[0], v[0], dst);
dst += 4;
- ++y;
+ y += 1;
u += (n & 1);
v += (n & 1);
- ++n;
}
}
static void YuvToBgraRow(const uint8_t* y, const uint8_t* u, const uint8_t* v,
uint8_t* dst, int len) {
+ const __m128i kAlpha = _mm_set1_epi16(255);
int n;
- for (n = 0; n + 2 <= len; n += 2) {
- const __m128i uv_0 = LoadUVPart(u[0], v[0]);
- const __m128i tmp0_1 = GetRGBA32bWithUV(y[0], uv_0);
- const __m128i tmp0_2 = GetRGBA32bWithUV(y[1], uv_0);
- const __m128i tmp1_1 = _mm_shuffle_epi32(tmp0_1, _MM_SHUFFLE(3, 0, 1, 2));
- const __m128i tmp1_2 = _mm_shuffle_epi32(tmp0_2, _MM_SHUFFLE(3, 0, 1, 2));
- const __m128i tmp2_1 = _mm_packs_epi32(tmp1_1, tmp1_2);
- const __m128i tmp3 = _mm_packus_epi16(tmp2_1, tmp2_1);
- _mm_storel_epi64((__m128i*)dst, tmp3);
- dst += 4 * 2;
- y += 2;
- ++u;
- ++v;
+ for (n = 0; n + 8 <= len; n += 8, dst += 32) {
+ __m128i R, G, B;
+ YUV420ToRGB(y, u, v, &R, &G, &B);
+ PackAndStore4(&B, &G, &R, &kAlpha, dst);
+ y += 8;
+ u += 4;
+ v += 4;
}
- // Finish off
- if (len & 1) {
+ for (; n < len; ++n) { // Finish off
VP8YuvToBgra(y[0], u[0], v[0], dst);
+ dst += 4;
+ y += 1;
+ u += (n & 1);
+ v += (n & 1);
}
}
static void YuvToArgbRow(const uint8_t* y, const uint8_t* u, const uint8_t* v,
uint8_t* dst, int len) {
+ const __m128i kAlpha = _mm_set1_epi16(255);
int n;
- for (n = 0; n + 2 <= len; n += 2) {
- const __m128i uv_0 = LoadUVPart(u[0], v[0]);
- const __m128i tmp0_1 = GetRGBA32bWithUV(y[0], uv_0);
- const __m128i tmp0_2 = GetRGBA32bWithUV(y[1], uv_0);
- const __m128i tmp1_1 = _mm_shuffle_epi32(tmp0_1, _MM_SHUFFLE(2, 1, 0, 3));
- const __m128i tmp1_2 = _mm_shuffle_epi32(tmp0_2, _MM_SHUFFLE(2, 1, 0, 3));
- const __m128i tmp2_1 = _mm_packs_epi32(tmp1_1, tmp1_2);
- const __m128i tmp3 = _mm_packus_epi16(tmp2_1, tmp2_1);
- _mm_storel_epi64((__m128i*)dst, tmp3);
- dst += 4 * 2;
- y += 2;
- ++u;
- ++v;
+ for (n = 0; n + 8 <= len; n += 8, dst += 32) {
+ __m128i R, G, B;
+ YUV420ToRGB(y, u, v, &R, &G, &B);
+ PackAndStore4(&kAlpha, &R, &G, &B, dst);
+ y += 8;
+ u += 4;
+ v += 4;
}
- // Finish off
- if (len & 1) {
+ for (; n < len; ++n) { // Finish off
VP8YuvToArgb(y[0], u[0], v[0], dst);
+ dst += 4;
+ y += 1;
+ u += (n & 1);
+ v += (n & 1);
}
}
static void YuvToRgbRow(const uint8_t* y, const uint8_t* u, const uint8_t* v,
uint8_t* dst, int len) {
int n;
- for (n = 0; n + 2 < len; ++n) { // we directly stomp the *dst memory
- YuvToRgbSSE2(y[0], u[0], v[0], dst); // stomps 8 bytes
+ for (n = 0; n + 32 <= len; n += 32, dst += 32 * 3) {
+ __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
+ __m128i rgb[6];
+
+ YUV420ToRGB(y + 0, u + 0, v + 0, &R0, &G0, &B0);
+ YUV420ToRGB(y + 8, u + 4, v + 4, &R1, &G1, &B1);
+ YUV420ToRGB(y + 16, u + 8, v + 8, &R2, &G2, &B2);
+ YUV420ToRGB(y + 24, u + 12, v + 12, &R3, &G3, &B3);
+
+ // Cast to 8b and store as RRRRGGGGBBBB.
+ rgb[0] = _mm_packus_epi16(R0, R1);
+ rgb[1] = _mm_packus_epi16(R2, R3);
+ rgb[2] = _mm_packus_epi16(G0, G1);
+ rgb[3] = _mm_packus_epi16(G2, G3);
+ rgb[4] = _mm_packus_epi16(B0, B1);
+ rgb[5] = _mm_packus_epi16(B2, B3);
+
+ // Pack as RGBRGBRGBRGB.
+ PlanarTo24b(rgb, dst);
+
+ y += 32;
+ u += 16;
+ v += 16;
+ }
+ for (; n < len; ++n) { // Finish off
+ VP8YuvToRgb(y[0], u[0], v[0], dst);
dst += 3;
- ++y;
+ y += 1;
u += (n & 1);
v += (n & 1);
}
- VP8YuvToRgb(y[0], u[0], v[0], dst);
- if (len > 1) {
- VP8YuvToRgb(y[1], u[n & 1], v[n & 1], dst + 3);
- }
}
static void YuvToBgrRow(const uint8_t* y, const uint8_t* u, const uint8_t* v,
uint8_t* dst, int len) {
int n;
- for (n = 0; n + 2 < len; ++n) { // we directly stomp the *dst memory
- YuvToBgrSSE2(y[0], u[0], v[0], dst); // stomps 8 bytes
+ for (n = 0; n + 32 <= len; n += 32, dst += 32 * 3) {
+ __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3;
+ __m128i bgr[6];
+
+ YUV420ToRGB(y + 0, u + 0, v + 0, &R0, &G0, &B0);
+ YUV420ToRGB(y + 8, u + 4, v + 4, &R1, &G1, &B1);
+ YUV420ToRGB(y + 16, u + 8, v + 8, &R2, &G2, &B2);
+ YUV420ToRGB(y + 24, u + 12, v + 12, &R3, &G3, &B3);
+
+ // Cast to 8b and store as BBBBGGGGRRRR.
+ bgr[0] = _mm_packus_epi16(B0, B1);
+ bgr[1] = _mm_packus_epi16(B2, B3);
+ bgr[2] = _mm_packus_epi16(G0, G1);
+ bgr[3] = _mm_packus_epi16(G2, G3);
+ bgr[4] = _mm_packus_epi16(R0, R1);
+ bgr[5] = _mm_packus_epi16(R2, R3);
+
+ // Pack as BGRBGRBGRBGR.
+ PlanarTo24b(bgr, dst);
+
+ y += 32;
+ u += 16;
+ v += 16;
+ }
+ for (; n < len; ++n) { // Finish off
+ VP8YuvToBgr(y[0], u[0], v[0], dst);
dst += 3;
- ++y;
+ y += 1;
u += (n & 1);
v += (n & 1);
}
- VP8YuvToBgr(y[0], u[0], v[0], dst + 0);
- if (len > 1) {
- VP8YuvToBgr(y[1], u[n & 1], v[n & 1], dst + 3);
- }
}
//------------------------------------------------------------------------------
@@ -316,52 +465,36 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitSamplersSSE2(void) {
// Store either 16b-words into *dst
#define STORE_16(V, dst) _mm_storeu_si128((__m128i*)(dst), (V))
-// Convert 8 packed RGB or BGR samples to r[], g[], b[]
+// Function that inserts a value of the second half of the in buffer in between
+// every two char of the first half.
+static WEBP_INLINE void RGB24PackedToPlanarHelper(
+ const __m128i* const in /*in[6]*/, __m128i* const out /*out[6]*/) {
+ out[0] = _mm_unpacklo_epi8(in[0], in[3]);
+ out[1] = _mm_unpackhi_epi8(in[0], in[3]);
+ out[2] = _mm_unpacklo_epi8(in[1], in[4]);
+ out[3] = _mm_unpackhi_epi8(in[1], in[4]);
+ out[4] = _mm_unpacklo_epi8(in[2], in[5]);
+ out[5] = _mm_unpackhi_epi8(in[2], in[5]);
+}
+
+// Unpack the 8b input rgbrgbrgbrgb ... as contiguous registers:
+// rrrr... rrrr... gggg... gggg... bbbb... bbbb....
+// Similar to PlanarTo24bHelper(), but in reverse order.
static WEBP_INLINE void RGB24PackedToPlanar(const uint8_t* const rgb,
- __m128i* const r,
- __m128i* const g,
- __m128i* const b,
- int input_is_bgr) {
- const __m128i zero = _mm_setzero_si128();
- // in0: r0 g0 b0 r1 | g1 b1 r2 g2 | b2 r3 g3 b3 | r4 g4 b4 r5
- // in1: b2 r3 g3 b3 | r4 g4 b4 r5 | g5 b5 r6 g6 | b6 r7 g7 b7
- const __m128i in0 = LOAD_16(rgb + 0);
- const __m128i in1 = LOAD_16(rgb + 8);
- // A0: | r2 g2 b2 r3 | g3 b3 r4 g4 | b4 r5 ...
- // A1: ... b2 r3 | g3 b3 r4 g4 | b4 r5 g5 b5 |
- const __m128i A0 = _mm_srli_si128(in0, 6);
- const __m128i A1 = _mm_slli_si128(in1, 6);
- // B0: r0 r2 g0 g2 | b0 b2 r1 r3 | g1 g3 b1 b3 | r2 r4 b2 b4
- // B1: g3 g5 b3 b5 | r4 r6 g4 g6 | b4 b6 r5 r7 | g5 g7 b5 b7
- const __m128i B0 = _mm_unpacklo_epi8(in0, A0);
- const __m128i B1 = _mm_unpackhi_epi8(A1, in1);
- // C0: r1 r3 g1 g3 | b1 b3 r2 r4 | b2 b4 ...
- // C1: ... g3 g5 | b3 b5 r4 r6 | g4 g6 b4 b6
- const __m128i C0 = _mm_srli_si128(B0, 6);
- const __m128i C1 = _mm_slli_si128(B1, 6);
- // D0: r0 r1 r2 r3 | g0 g1 g2 g3 | b0 b1 b2 b3 | r1 r2 r3 r4
- // D1: b3 b4 b5 b6 | r4 r5 r6 r7 | g4 g5 g6 g7 | b4 b5 b6 b7 |
- const __m128i D0 = _mm_unpacklo_epi8(B0, C0);
- const __m128i D1 = _mm_unpackhi_epi8(C1, B1);
- // r4 r5 r6 r7 | g4 g5 g6 g7 | b4 b5 b6 b7 | 0
- const __m128i D2 = _mm_srli_si128(D1, 4);
- // r0 r1 r2 r3 | r4 r5 r6 r7 | g0 g1 g2 g3 | g4 g5 g6 g7
- const __m128i E0 = _mm_unpacklo_epi32(D0, D2);
- // b0 b1 b2 b3 | b4 b5 b6 b7 | r1 r2 r3 r4 | 0
- const __m128i E1 = _mm_unpackhi_epi32(D0, D2);
- // g0 g1 g2 g3 | g4 g5 g6 g7 | 0
- const __m128i E2 = _mm_srli_si128(E0, 8);
- const __m128i F0 = _mm_unpacklo_epi8(E0, zero); // -> R
- const __m128i F1 = _mm_unpacklo_epi8(E1, zero); // -> B
- const __m128i F2 = _mm_unpacklo_epi8(E2, zero); // -> G
- *g = F2;
- if (input_is_bgr) {
- *r = F1;
- *b = F0;
- } else {
- *r = F0;
- *b = F1;
- }
+ __m128i* const out /*out[6]*/) {
+ __m128i tmp[6];
+ tmp[0] = _mm_loadu_si128((const __m128i*)(rgb + 0));
+ tmp[1] = _mm_loadu_si128((const __m128i*)(rgb + 16));
+ tmp[2] = _mm_loadu_si128((const __m128i*)(rgb + 32));
+ tmp[3] = _mm_loadu_si128((const __m128i*)(rgb + 48));
+ tmp[4] = _mm_loadu_si128((const __m128i*)(rgb + 64));
+ tmp[5] = _mm_loadu_si128((const __m128i*)(rgb + 80));
+
+ RGB24PackedToPlanarHelper(tmp, out);
+ RGB24PackedToPlanarHelper(out, tmp);
+ RGB24PackedToPlanarHelper(tmp, out);
+ RGB24PackedToPlanarHelper(out, tmp);
+ RGB24PackedToPlanarHelper(tmp, out);
}
// Convert 8 packed ARGB to r[], g[], b[]
@@ -445,15 +578,33 @@ static WEBP_INLINE void ConvertRGBToUV(const __m128i* const R,
#undef TRANSFORM
static void ConvertRGB24ToY(const uint8_t* rgb, uint8_t* y, int width) {
- const int max_width = width & ~15;
+ const int max_width = width & ~31;
int i;
- for (i = 0; i < max_width; i += 16, rgb += 3 * 16) {
- __m128i r, g, b, Y0, Y1;
- RGB24PackedToPlanar(rgb + 0 * 8, &r, &g, &b, 0);
- ConvertRGBToY(&r, &g, &b, &Y0);
- RGB24PackedToPlanar(rgb + 3 * 8, &r, &g, &b, 0);
- ConvertRGBToY(&r, &g, &b, &Y1);
- STORE_16(_mm_packus_epi16(Y0, Y1), y + i);
+ for (i = 0; i < max_width; rgb += 3 * 16 * 2) {
+ __m128i rgb_plane[6];
+ int j;
+
+ RGB24PackedToPlanar(rgb, rgb_plane);
+
+ for (j = 0; j < 2; ++j, i += 16) {
+ const __m128i zero = _mm_setzero_si128();
+ __m128i r, g, b, Y0, Y1;
+
+ // Convert to 16-bit Y.
+ r = _mm_unpacklo_epi8(rgb_plane[0 + j], zero);
+ g = _mm_unpacklo_epi8(rgb_plane[2 + j], zero);
+ b = _mm_unpacklo_epi8(rgb_plane[4 + j], zero);
+ ConvertRGBToY(&r, &g, &b, &Y0);
+
+ // Convert to 16-bit Y.
+ r = _mm_unpackhi_epi8(rgb_plane[0 + j], zero);
+ g = _mm_unpackhi_epi8(rgb_plane[2 + j], zero);
+ b = _mm_unpackhi_epi8(rgb_plane[4 + j], zero);
+ ConvertRGBToY(&r, &g, &b, &Y1);
+
+ // Cast to 8-bit and store.
+ STORE_16(_mm_packus_epi16(Y0, Y1), y + i);
+ }
}
for (; i < width; ++i, rgb += 3) { // left-over
y[i] = VP8RGBToY(rgb[0], rgb[1], rgb[2], YUV_HALF);
@@ -461,15 +612,33 @@ static void ConvertRGB24ToY(const uint8_t* rgb, uint8_t* y, int width) {
}
static void ConvertBGR24ToY(const uint8_t* bgr, uint8_t* y, int width) {
+ const int max_width = width & ~31;
int i;
- const int max_width = width & ~15;
- for (i = 0; i < max_width; i += 16, bgr += 3 * 16) {
- __m128i r, g, b, Y0, Y1;
- RGB24PackedToPlanar(bgr + 0 * 8, &r, &g, &b, 1);
- ConvertRGBToY(&r, &g, &b, &Y0);
- RGB24PackedToPlanar(bgr + 3 * 8, &r, &g, &b, 1);
- ConvertRGBToY(&r, &g, &b, &Y1);
- STORE_16(_mm_packus_epi16(Y0, Y1), y + i);
+ for (i = 0; i < max_width; bgr += 3 * 16 * 2) {
+ __m128i bgr_plane[6];
+ int j;
+
+ RGB24PackedToPlanar(bgr, bgr_plane);
+
+ for (j = 0; j < 2; ++j, i += 16) {
+ const __m128i zero = _mm_setzero_si128();
+ __m128i r, g, b, Y0, Y1;
+
+ // Convert to 16-bit Y.
+ b = _mm_unpacklo_epi8(bgr_plane[0 + j], zero);
+ g = _mm_unpacklo_epi8(bgr_plane[2 + j], zero);
+ r = _mm_unpacklo_epi8(bgr_plane[4 + j], zero);
+ ConvertRGBToY(&r, &g, &b, &Y0);
+
+ // Convert to 16-bit Y.
+ b = _mm_unpackhi_epi8(bgr_plane[0 + j], zero);
+ g = _mm_unpackhi_epi8(bgr_plane[2 + j], zero);
+ r = _mm_unpackhi_epi8(bgr_plane[4 + j], zero);
+ ConvertRGBToY(&r, &g, &b, &Y1);
+
+ // Cast to 8-bit and store.
+ STORE_16(_mm_packus_epi16(Y0, Y1), y + i);
+ }
}
for (; i < width; ++i, bgr += 3) { // left-over
y[i] = VP8RGBToY(bgr[2], bgr[1], bgr[0], YUV_HALF);
generated by cgit v1.2.3 (git 2.25.1) at 2024-12-26 17:45:38 +0000