Lines Matching refs:__m128i
47 const __m128i *s = reinterpret_cast<const __m128i*>(src); in S32_Blend_BlitRow32_SSE2()
48 __m128i *d = reinterpret_cast<__m128i*>(dst); in S32_Blend_BlitRow32_SSE2()
49 __m128i rb_mask = _mm_set1_epi32(0x00FF00FF); in S32_Blend_BlitRow32_SSE2()
50 __m128i src_scale_wide = _mm_set1_epi16(src_scale); in S32_Blend_BlitRow32_SSE2()
51 __m128i dst_scale_wide = _mm_set1_epi16(dst_scale); in S32_Blend_BlitRow32_SSE2()
54 __m128i src_pixel = _mm_loadu_si128(s); in S32_Blend_BlitRow32_SSE2()
55 __m128i dst_pixel = _mm_load_si128(d); in S32_Blend_BlitRow32_SSE2()
58 __m128i dst_rb = _mm_and_si128(rb_mask, dst_pixel); in S32_Blend_BlitRow32_SSE2()
59 __m128i src_rb = _mm_and_si128(rb_mask, src_pixel); in S32_Blend_BlitRow32_SSE2()
62 __m128i dst_ag = _mm_srli_epi16(dst_pixel, 8); in S32_Blend_BlitRow32_SSE2()
63 __m128i src_ag = _mm_srli_epi16(src_pixel, 8); in S32_Blend_BlitRow32_SSE2()
82 __m128i result = _mm_add_epi8(src_pixel, dst_pixel); in S32_Blend_BlitRow32_SSE2()
117 const __m128i *s = reinterpret_cast<const __m128i*>(src); in S32A_Opaque_BlitRow32_SSE2()
118 __m128i *d = reinterpret_cast<__m128i*>(dst); in S32A_Opaque_BlitRow32_SSE2()
120 __m128i rb_mask = _mm_set1_epi32(0x00FF00FF); in S32A_Opaque_BlitRow32_SSE2()
121 __m128i c_128 = _mm_set1_epi16(128); // 8 copies of 128 (16-bit) in S32A_Opaque_BlitRow32_SSE2()
122 __m128i c_255 = _mm_set1_epi16(255); // 8 copies of 255 (16-bit) in S32A_Opaque_BlitRow32_SSE2()
125 __m128i src_pixel = _mm_loadu_si128(s); in S32A_Opaque_BlitRow32_SSE2()
126 __m128i dst_pixel = _mm_load_si128(d); in S32A_Opaque_BlitRow32_SSE2()
128 __m128i dst_rb = _mm_and_si128(rb_mask, dst_pixel); in S32A_Opaque_BlitRow32_SSE2()
129 __m128i dst_ag = _mm_srli_epi16(dst_pixel, 8); in S32A_Opaque_BlitRow32_SSE2()
131 __m128i alpha = _mm_srli_epi32(src_pixel, 24); in S32A_Opaque_BlitRow32_SSE2()
145 __m128i dst_rb_low = _mm_srli_epi16(dst_rb, 8); in S32A_Opaque_BlitRow32_SSE2()
146 __m128i dst_ag_low = _mm_srli_epi16(dst_ag, 8); in S32A_Opaque_BlitRow32_SSE2()
162 __m128i result = _mm_add_epi8(src_pixel, dst_pixel); in S32A_Opaque_BlitRow32_SSE2()
169 __m128i rb_mask = _mm_set1_epi32(0x00FF00FF); in S32A_Opaque_BlitRow32_SSE2()
170 __m128i c_256 = _mm_set1_epi16(0x0100); // 8 copies of 256 (16-bit) in S32A_Opaque_BlitRow32_SSE2()
173 __m128i src_pixel = _mm_loadu_si128(s); in S32A_Opaque_BlitRow32_SSE2()
174 __m128i dst_pixel = _mm_load_si128(d); in S32A_Opaque_BlitRow32_SSE2()
176 __m128i dst_rb = _mm_and_si128(rb_mask, dst_pixel); in S32A_Opaque_BlitRow32_SSE2()
177 __m128i dst_ag = _mm_srli_epi16(dst_pixel, 8); in S32A_Opaque_BlitRow32_SSE2()
180 __m128i alpha = _mm_srli_epi16(src_pixel, 8); in S32A_Opaque_BlitRow32_SSE2()
206 __m128i result = _mm_add_epi8(src_pixel, dst_pixel); in S32A_Opaque_BlitRow32_SSE2()
243 const __m128i *s = reinterpret_cast<const __m128i*>(src); in S32A_Blend_BlitRow32_SSE2()
244 __m128i *d = reinterpret_cast<__m128i*>(dst); in S32A_Blend_BlitRow32_SSE2()
245 __m128i src_scale_wide = _mm_set1_epi16(src_scale); in S32A_Blend_BlitRow32_SSE2()
246 __m128i rb_mask = _mm_set1_epi32(0x00FF00FF); in S32A_Blend_BlitRow32_SSE2()
247 __m128i c_256 = _mm_set1_epi16(256); // 8 copies of 256 (16-bit) in S32A_Blend_BlitRow32_SSE2()
250 __m128i src_pixel = _mm_loadu_si128(s); in S32A_Blend_BlitRow32_SSE2()
251 __m128i dst_pixel = _mm_load_si128(d); in S32A_Blend_BlitRow32_SSE2()
254 __m128i dst_rb = _mm_and_si128(rb_mask, dst_pixel); in S32A_Blend_BlitRow32_SSE2()
255 __m128i src_rb = _mm_and_si128(rb_mask, src_pixel); in S32A_Blend_BlitRow32_SSE2()
258 __m128i dst_ag = _mm_srli_epi16(dst_pixel, 8); in S32A_Blend_BlitRow32_SSE2()
259 __m128i src_ag = _mm_srli_epi16(src_pixel, 8); in S32A_Blend_BlitRow32_SSE2()
262 __m128i dst_alpha = _mm_shufflehi_epi16(src_ag, 0xF5); in S32A_Blend_BlitRow32_SSE2()
297 __m128i result = _mm_add_epi8(src_pixel, dst_pixel); in S32A_Blend_BlitRow32_SSE2()
346 const __m128i *s = reinterpret_cast<const __m128i*>(src); in Color32_SSE2()
347 __m128i *d = reinterpret_cast<__m128i*>(dst); in Color32_SSE2()
348 __m128i rb_mask = _mm_set1_epi32(0x00FF00FF); in Color32_SSE2()
349 __m128i src_scale_wide = _mm_set1_epi16(scale); in Color32_SSE2()
350 __m128i color_wide = _mm_set1_epi32(color); in Color32_SSE2()
353 __m128i src_pixel = _mm_loadu_si128(s); in Color32_SSE2()
356 __m128i src_rb = _mm_and_si128(rb_mask, src_pixel); in Color32_SSE2()
359 __m128i src_ag = _mm_srli_epi16(src_pixel, 8); in Color32_SSE2()
373 __m128i result = _mm_add_epi8(color_wide, src_pixel); in Color32_SSE2()
412 __m128i *d = reinterpret_cast<__m128i*>(dst); in SkARGB32_BlitMask_SSE2()
413 __m128i rb_mask = _mm_set1_epi32(0x00FF00FF); in SkARGB32_BlitMask_SSE2()
414 __m128i c_256 = _mm_set1_epi16(256); in SkARGB32_BlitMask_SSE2()
415 __m128i c_1 = _mm_set1_epi16(1); in SkARGB32_BlitMask_SSE2()
416 __m128i src_pixel = _mm_set1_epi32(color); in SkARGB32_BlitMask_SSE2()
419 __m128i dst_pixel = _mm_load_si128(d); in SkARGB32_BlitMask_SSE2()
422 __m128i src_scale_wide = _mm_set_epi8(0, *(mask+3),\ in SkARGB32_BlitMask_SSE2()
432 __m128i dst_rb = _mm_and_si128(rb_mask, dst_pixel); in SkARGB32_BlitMask_SSE2()
433 __m128i src_rb = _mm_and_si128(rb_mask, src_pixel); in SkARGB32_BlitMask_SSE2()
436 __m128i dst_ag = _mm_srli_epi16(dst_pixel, 8); in SkARGB32_BlitMask_SSE2()
437 __m128i src_ag = _mm_srli_epi16(src_pixel, 8); in SkARGB32_BlitMask_SSE2()
440 __m128i dst_alpha = _mm_shufflehi_epi16(src_ag, 0xF5); in SkARGB32_BlitMask_SSE2()
470 __m128i tmp_src_pixel = _mm_or_si128(src_rb, src_ag); in SkARGB32_BlitMask_SSE2()
473 __m128i result = _mm_add_epi8(tmp_src_pixel, dst_pixel); in SkARGB32_BlitMask_SSE2()