diff options
Diffstat (limited to 'src/dsp/upsampling_sse2.c')
| -rw-r--r-- | src/dsp/upsampling_sse2.c | 38 |
1 files changed, 16 insertions, 22 deletions
diff --git a/src/dsp/upsampling_sse2.c b/src/dsp/upsampling_sse2.c index ba075d11..8cb275a0 100644 --- a/src/dsp/upsampling_sse2.c +++ b/src/dsp/upsampling_sse2.c @@ -11,10 +11,6 @@ #include "./dsp.h" -#if defined(__cplusplus) || defined(c_plusplus) -extern "C" { -#endif - #if defined(WEBP_USE_SSE2) #include <assert.h> @@ -22,6 +18,10 @@ extern "C" { #include <string.h> #include "./yuv.h" +#if defined(__cplusplus) || defined(c_plusplus) +extern "C" { +#endif + #ifdef FANCY_UPSAMPLING // We compute (9*a + 3*b + 3*c + d + 8) / 16 as follows @@ -51,12 +51,12 @@ extern "C" { // pack and store two alterning pixel rows #define PACK_AND_STORE(a, b, da, db, out) do { \ - const __m128i t_a = _mm_avg_epu8(a, da); /* (9a + 3b + 3c + d + 8) / 16 */ \ - const __m128i t_b = _mm_avg_epu8(b, db); /* (3a + 9b + c + 3d + 8) / 16 */ \ - const __m128i t_1 = _mm_unpacklo_epi8(t_a, t_b); \ - const __m128i t_2 = _mm_unpackhi_epi8(t_a, t_b); \ - _mm_store_si128(((__m128i*)(out)) + 0, t_1); \ - _mm_store_si128(((__m128i*)(out)) + 1, t_2); \ + const __m128i ta = _mm_avg_epu8(a, da); /* (9a + 3b + 3c + d + 8) / 16 */ \ + const __m128i tb = _mm_avg_epu8(b, db); /* (3a + 9b + c + 3d + 8) / 16 */ \ + const __m128i t1 = _mm_unpacklo_epi8(ta, tb); \ + const __m128i t2 = _mm_unpackhi_epi8(ta, tb); \ + _mm_store_si128(((__m128i*)(out)) + 0, t1); \ + _mm_store_si128(((__m128i*)(out)) + 1, t2); \ } while (0) // Loads 17 pixels each from rows r1 and r2 and generates 32 pixels. @@ -128,7 +128,7 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \ const uint8_t* top_u, const uint8_t* top_v, \ const uint8_t* cur_u, const uint8_t* cur_v, \ uint8_t* top_dst, uint8_t* bottom_dst, int len) { \ - int block; \ + int b; \ /* 16 byte aligned array to cache reconstructed u and v */ \ uint8_t uv_buf[4 * 32 + 15]; \ uint8_t* const r_uv = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15); \ @@ -154,11 +154,11 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \ FUNC(bottom_y[0], u0, v0, bottom_dst); \ } \ \ - for (block = 0; block < num_blocks; ++block) { \ + for (b = 0; b < num_blocks; ++b) { \ UPSAMPLE_32PIXELS(top_u, cur_u, r_uv + 0 * 32); \ UPSAMPLE_32PIXELS(top_v, cur_v, r_uv + 1 * 32); \ CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y, r_uv, top_dst, bottom_dst, \ - 32 * block + 1, 32) \ + 32 * b + 1, 32) \ top_u += 16; \ cur_u += 16; \ top_v += 16; \ @@ -184,32 +184,26 @@ SSE2_UPSAMPLE_FUNC(UpsampleBgraLinePairSSE2, VP8YuvToBgra, 4) #undef CONVERT2RGB #undef SSE2_UPSAMPLE_FUNC -#endif // FANCY_UPSAMPLING - -#endif // WEBP_USE_SSE2 - //------------------------------------------------------------------------------ extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */]; void WebPInitUpsamplersSSE2(void) { -#if defined(WEBP_USE_SSE2) WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePairSSE2; WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePairSSE2; WebPUpsamplers[MODE_BGR] = UpsampleBgrLinePairSSE2; WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePairSSE2; -#endif // WEBP_USE_SSE2 } void WebPInitPremultiplySSE2(void) { -#if defined(WEBP_USE_SSE2) WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePairSSE2; WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePairSSE2; -#endif // WEBP_USE_SSE2 } +#endif // FANCY_UPSAMPLING + #if defined(__cplusplus) || defined(c_plusplus) } // extern "C" #endif - +#endif // WEBP_USE_SSE2 |