Home
last modified time | relevance | path

Searched refs:__m128i (Results 1 – 25 of 162) sorted by relevance

1234567

/third_party/skia/third_party/externals/libwebp/src/dsp/
Dlossless_enc_sse2.c33 const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb in SubtractGreenFromBlueAndRed_SSE2()
34 const __m128i A = _mm_srli_epi16(in, 8); // 0 a 0 g in SubtractGreenFromBlueAndRed_SSE2()
35 const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0)); in SubtractGreenFromBlueAndRed_SSE2()
36 const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // 0g0g in SubtractGreenFromBlueAndRed_SSE2()
37 const __m128i out = _mm_sub_epi8(in, C); in SubtractGreenFromBlueAndRed_SSE2()
38 _mm_storeu_si128((__m128i*)&argb_data[i], out); in SubtractGreenFromBlueAndRed_SSE2()
54 const __m128i mults_rb = MK_CST_16(CST_5b(m->green_to_red_), in TransformColor_SSE2()
56 const __m128i mults_b2 = MK_CST_16(CST_5b(m->red_to_blue_), 0); in TransformColor_SSE2()
57 const __m128i mask_ag = _mm_set1_epi32(0xff00ff00); // alpha-green masks in TransformColor_SSE2()
58 const __m128i mask_rb = _mm_set1_epi32(0x00ff00ff); // red-blue masks in TransformColor_SSE2()
[all …]
Denc_sse2.c47 const __m128i k1 = _mm_set1_epi16(20091); in ITransform_SSE2()
48 const __m128i k2 = _mm_set1_epi16(-30068); in ITransform_SSE2()
49 __m128i T0, T1, T2, T3; in ITransform_SSE2()
55 __m128i in0, in1, in2, in3; in ITransform_SSE2()
57 in0 = _mm_loadl_epi64((const __m128i*)&in[0]); in ITransform_SSE2()
58 in1 = _mm_loadl_epi64((const __m128i*)&in[4]); in ITransform_SSE2()
59 in2 = _mm_loadl_epi64((const __m128i*)&in[8]); in ITransform_SSE2()
60 in3 = _mm_loadl_epi64((const __m128i*)&in[12]); in ITransform_SSE2()
66 const __m128i inB0 = _mm_loadl_epi64((const __m128i*)&in[16]); in ITransform_SSE2()
67 const __m128i inB1 = _mm_loadl_epi64((const __m128i*)&in[20]); in ITransform_SSE2()
[all …]
Dlossless_sse2.c29 const __m128i zero = _mm_setzero_si128(); in ClampedAddSubtractFull_SSE2()
30 const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero); in ClampedAddSubtractFull_SSE2()
31 const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero); in ClampedAddSubtractFull_SSE2()
32 const __m128i C2 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero); in ClampedAddSubtractFull_SSE2()
33 const __m128i V1 = _mm_add_epi16(C0, C1); in ClampedAddSubtractFull_SSE2()
34 const __m128i V2 = _mm_sub_epi16(V1, C2); in ClampedAddSubtractFull_SSE2()
35 const __m128i b = _mm_packus_epi16(V2, V2); in ClampedAddSubtractFull_SSE2()
43 const __m128i zero = _mm_setzero_si128(); in ClampedAddSubtractHalf_SSE2()
44 const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero); in ClampedAddSubtractHalf_SSE2()
45 const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero); in ClampedAddSubtractHalf_SSE2()
[all …]
Dalpha_processing_sse2.c28 const __m128i zero = _mm_setzero_si128(); in DispatchAlpha_SSE2()
29 const __m128i rgb_mask = _mm_set1_epi32(0xffffff00u); // to preserve RGB in DispatchAlpha_SSE2()
30 const __m128i all_0xff = _mm_set_epi32(0, 0, ~0u, ~0u); in DispatchAlpha_SSE2()
31 __m128i all_alphas = all_0xff; in DispatchAlpha_SSE2()
39 __m128i* out = (__m128i*)dst; in DispatchAlpha_SSE2()
42 const __m128i a0 = _mm_loadl_epi64((const __m128i*)&alpha[i]); in DispatchAlpha_SSE2()
43 const __m128i a1 = _mm_unpacklo_epi8(a0, zero); in DispatchAlpha_SSE2()
44 const __m128i a2_lo = _mm_unpacklo_epi16(a1, zero); in DispatchAlpha_SSE2()
45 const __m128i a2_hi = _mm_unpackhi_epi16(a1, zero); in DispatchAlpha_SSE2()
47 const __m128i b0_lo = _mm_loadu_si128(out + 0); in DispatchAlpha_SSE2()
[all …]
Denc_sse41.c29 const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH); in CollectHistogram_SSE41()
41 const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]); in CollectHistogram_SSE41()
42 const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]); in CollectHistogram_SSE41()
44 const __m128i abs0 = _mm_abs_epi16(out0); in CollectHistogram_SSE41()
45 const __m128i abs1 = _mm_abs_epi16(out1); in CollectHistogram_SSE41()
46 const __m128i v0 = _mm_srai_epi16(abs0, 3); in CollectHistogram_SSE41()
47 const __m128i v1 = _mm_srai_epi16(abs1, 3); in CollectHistogram_SSE41()
49 const __m128i bin0 = _mm_min_epi16(v0, max_coeff_thresh); in CollectHistogram_SSE41()
50 const __m128i bin1 = _mm_min_epi16(v1, max_coeff_thresh); in CollectHistogram_SSE41()
52 _mm_storeu_si128((__m128i*)&out[0], bin0); in CollectHistogram_SSE41()
[all …]
Ddec_sse2.c50 const __m128i k1 = _mm_set1_epi16(20091); in Transform_SSE2()
51 const __m128i k2 = _mm_set1_epi16(-30068); in Transform_SSE2()
52 __m128i T0, T1, T2, T3; in Transform_SSE2()
57 __m128i in0, in1, in2, in3; in Transform_SSE2()
59 in0 = _mm_loadl_epi64((const __m128i*)&in[0]); in Transform_SSE2()
60 in1 = _mm_loadl_epi64((const __m128i*)&in[4]); in Transform_SSE2()
61 in2 = _mm_loadl_epi64((const __m128i*)&in[8]); in Transform_SSE2()
62 in3 = _mm_loadl_epi64((const __m128i*)&in[12]); in Transform_SSE2()
68 const __m128i inB0 = _mm_loadl_epi64((const __m128i*)&in[16]); in Transform_SSE2()
69 const __m128i inB1 = _mm_loadl_epi64((const __m128i*)&in[20]); in Transform_SSE2()
[all …]
Dyuv_sse2.c29 static void ConvertYUV444ToRGB_SSE2(const __m128i* const Y0, in ConvertYUV444ToRGB_SSE2()
30 const __m128i* const U0, in ConvertYUV444ToRGB_SSE2()
31 const __m128i* const V0, in ConvertYUV444ToRGB_SSE2()
32 __m128i* const R, in ConvertYUV444ToRGB_SSE2()
33 __m128i* const G, in ConvertYUV444ToRGB_SSE2()
34 __m128i* const B) { in ConvertYUV444ToRGB_SSE2()
35 const __m128i k19077 = _mm_set1_epi16(19077); in ConvertYUV444ToRGB_SSE2()
36 const __m128i k26149 = _mm_set1_epi16(26149); in ConvertYUV444ToRGB_SSE2()
37 const __m128i k14234 = _mm_set1_epi16(14234); in ConvertYUV444ToRGB_SSE2()
39 const __m128i k33050 = _mm_set1_epi16((short)33050); in ConvertYUV444ToRGB_SSE2()
[all …]
Drescaler_sse2.c31 static void LoadTwoPixels_SSE2(const uint8_t* const src, __m128i* out) { in LoadTwoPixels_SSE2()
32 const __m128i zero = _mm_setzero_si128(); in LoadTwoPixels_SSE2()
33 const __m128i A = _mm_loadl_epi64((const __m128i*)(src)); // ABCDEFGH in LoadTwoPixels_SSE2()
34 const __m128i B = _mm_unpacklo_epi8(A, zero); // A0B0C0D0E0F0G0H0 in LoadTwoPixels_SSE2()
35 const __m128i C = _mm_srli_si128(B, 8); // E0F0G0H0 in LoadTwoPixels_SSE2()
40 static void LoadEightPixels_SSE2(const uint8_t* const src, __m128i* out) { in LoadEightPixels_SSE2()
41 const __m128i zero = _mm_setzero_si128(); in LoadEightPixels_SSE2()
42 const __m128i A = _mm_loadl_epi64((const __m128i*)(src)); // ABCDEFGH in LoadEightPixels_SSE2()
52 __m128i cur_pixels; in RescalerImportRowExpand_SSE2()
66 const __m128i mult = _mm_set1_epi32(((x_add - accum) << 16) | accum); in RescalerImportRowExpand_SSE2()
[all …]
Dyuv_sse41.c29 static void ConvertYUV444ToRGB_SSE41(const __m128i* const Y0, in ConvertYUV444ToRGB_SSE41()
30 const __m128i* const U0, in ConvertYUV444ToRGB_SSE41()
31 const __m128i* const V0, in ConvertYUV444ToRGB_SSE41()
32 __m128i* const R, in ConvertYUV444ToRGB_SSE41()
33 __m128i* const G, in ConvertYUV444ToRGB_SSE41()
34 __m128i* const B) { in ConvertYUV444ToRGB_SSE41()
35 const __m128i k19077 = _mm_set1_epi16(19077); in ConvertYUV444ToRGB_SSE41()
36 const __m128i k26149 = _mm_set1_epi16(26149); in ConvertYUV444ToRGB_SSE41()
37 const __m128i k14234 = _mm_set1_epi16(14234); in ConvertYUV444ToRGB_SSE41()
39 const __m128i k33050 = _mm_set1_epi16((short)33050); in ConvertYUV444ToRGB_SSE41()
[all …]
Dlossless_sse41.c28 const __m128i mults_rb = _mm_set1_epi32((uint32_t)CST(green_to_red_) << 16 | in TransformColorInverse_SSE41()
30 const __m128i mults_b2 = _mm_set1_epi32(CST(red_to_blue_)); in TransformColorInverse_SSE41()
32 const __m128i mask_ag = _mm_set1_epi32(0xff00ff00); in TransformColorInverse_SSE41()
33 const __m128i perm1 = _mm_setr_epi8(-1, 1, -1, 1, -1, 5, -1, 5, in TransformColorInverse_SSE41()
35 const __m128i perm2 = _mm_setr_epi8(-1, 2, -1, -1, -1, 6, -1, -1, in TransformColorInverse_SSE41()
39 const __m128i A = _mm_loadu_si128((const __m128i*)(src + i)); in TransformColorInverse_SSE41()
40 const __m128i B = _mm_shuffle_epi8(A, perm1); // argb -> g0g0 in TransformColorInverse_SSE41()
41 const __m128i C = _mm_mulhi_epi16(B, mults_rb); in TransformColorInverse_SSE41()
42 const __m128i D = _mm_add_epi8(A, C); in TransformColorInverse_SSE41()
43 const __m128i E = _mm_shuffle_epi8(D, perm2); in TransformColorInverse_SSE41()
[all …]
Dcommon_sse2.h30 static WEBP_INLINE void PrintReg(const __m128i r, const char* const name,
34 __m128i r;
59 static WEBP_INLINE int VP8HorizontalAdd8b(const __m128i* const a) { in VP8HorizontalAdd8b()
60 const __m128i zero = _mm_setzero_si128(); in VP8HorizontalAdd8b()
61 const __m128i sad8x2 = _mm_sad_epu8(*a, zero); in VP8HorizontalAdd8b()
63 const __m128i sum = _mm_add_epi32(sad8x2, _mm_shuffle_epi32(sad8x2, 2)); in VP8HorizontalAdd8b()
69 const __m128i* const in0, const __m128i* const in1, in VP8Transpose_2_4x4_16b()
70 const __m128i* const in2, const __m128i* const in3, __m128i* const out0, in VP8Transpose_2_4x4_16b()
71 __m128i* const out1, __m128i* const out2, __m128i* const out3) { in VP8Transpose_2_4x4_16b()
77 const __m128i transpose0_0 = _mm_unpacklo_epi16(*in0, *in1); in VP8Transpose_2_4x4_16b()
[all …]
/third_party/flutter/skia/third_party/externals/libwebp/src/dsp/
Dlossless_enc_sse2.c33 const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb in SubtractGreenFromBlueAndRed_SSE2()
34 const __m128i A = _mm_srli_epi16(in, 8); // 0 a 0 g in SubtractGreenFromBlueAndRed_SSE2()
35 const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0)); in SubtractGreenFromBlueAndRed_SSE2()
36 const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // 0g0g in SubtractGreenFromBlueAndRed_SSE2()
37 const __m128i out = _mm_sub_epi8(in, C); in SubtractGreenFromBlueAndRed_SSE2()
38 _mm_storeu_si128((__m128i*)&argb_data[i], out); in SubtractGreenFromBlueAndRed_SSE2()
54 const __m128i mults_rb = MK_CST_16(CST_5b(m->green_to_red_), in TransformColor_SSE2()
56 const __m128i mults_b2 = MK_CST_16(CST_5b(m->red_to_blue_), 0); in TransformColor_SSE2()
57 const __m128i mask_ag = _mm_set1_epi32(0xff00ff00); // alpha-green masks in TransformColor_SSE2()
58 const __m128i mask_rb = _mm_set1_epi32(0x00ff00ff); // red-blue masks in TransformColor_SSE2()
[all …]
Denc_sse2.c47 const __m128i k1 = _mm_set1_epi16(20091); in ITransform_SSE2()
48 const __m128i k2 = _mm_set1_epi16(-30068); in ITransform_SSE2()
49 __m128i T0, T1, T2, T3; in ITransform_SSE2()
55 __m128i in0, in1, in2, in3; in ITransform_SSE2()
57 in0 = _mm_loadl_epi64((const __m128i*)&in[0]); in ITransform_SSE2()
58 in1 = _mm_loadl_epi64((const __m128i*)&in[4]); in ITransform_SSE2()
59 in2 = _mm_loadl_epi64((const __m128i*)&in[8]); in ITransform_SSE2()
60 in3 = _mm_loadl_epi64((const __m128i*)&in[12]); in ITransform_SSE2()
66 const __m128i inB0 = _mm_loadl_epi64((const __m128i*)&in[16]); in ITransform_SSE2()
67 const __m128i inB1 = _mm_loadl_epi64((const __m128i*)&in[20]); in ITransform_SSE2()
[all …]
Dlossless_sse2.c30 const __m128i zero = _mm_setzero_si128(); in ClampedAddSubtractFull_SSE2()
31 const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero); in ClampedAddSubtractFull_SSE2()
32 const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero); in ClampedAddSubtractFull_SSE2()
33 const __m128i C2 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero); in ClampedAddSubtractFull_SSE2()
34 const __m128i V1 = _mm_add_epi16(C0, C1); in ClampedAddSubtractFull_SSE2()
35 const __m128i V2 = _mm_sub_epi16(V1, C2); in ClampedAddSubtractFull_SSE2()
36 const __m128i b = _mm_packus_epi16(V2, V2); in ClampedAddSubtractFull_SSE2()
44 const __m128i zero = _mm_setzero_si128(); in ClampedAddSubtractHalf_SSE2()
45 const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero); in ClampedAddSubtractHalf_SSE2()
46 const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero); in ClampedAddSubtractHalf_SSE2()
[all …]
Dalpha_processing_sse2.c28 const __m128i zero = _mm_setzero_si128(); in DispatchAlpha_SSE2()
29 const __m128i rgb_mask = _mm_set1_epi32(0xffffff00u); // to preserve RGB in DispatchAlpha_SSE2()
30 const __m128i all_0xff = _mm_set_epi32(0, 0, ~0u, ~0u); in DispatchAlpha_SSE2()
31 __m128i all_alphas = all_0xff; in DispatchAlpha_SSE2()
39 __m128i* out = (__m128i*)dst; in DispatchAlpha_SSE2()
42 const __m128i a0 = _mm_loadl_epi64((const __m128i*)&alpha[i]); in DispatchAlpha_SSE2()
43 const __m128i a1 = _mm_unpacklo_epi8(a0, zero); in DispatchAlpha_SSE2()
44 const __m128i a2_lo = _mm_unpacklo_epi16(a1, zero); in DispatchAlpha_SSE2()
45 const __m128i a2_hi = _mm_unpackhi_epi16(a1, zero); in DispatchAlpha_SSE2()
47 const __m128i b0_lo = _mm_loadu_si128(out + 0); in DispatchAlpha_SSE2()
[all …]
Denc_sse41.c29 const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH); in CollectHistogram_SSE41()
41 const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]); in CollectHistogram_SSE41()
42 const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]); in CollectHistogram_SSE41()
44 const __m128i abs0 = _mm_abs_epi16(out0); in CollectHistogram_SSE41()
45 const __m128i abs1 = _mm_abs_epi16(out1); in CollectHistogram_SSE41()
46 const __m128i v0 = _mm_srai_epi16(abs0, 3); in CollectHistogram_SSE41()
47 const __m128i v1 = _mm_srai_epi16(abs1, 3); in CollectHistogram_SSE41()
49 const __m128i bin0 = _mm_min_epi16(v0, max_coeff_thresh); in CollectHistogram_SSE41()
50 const __m128i bin1 = _mm_min_epi16(v1, max_coeff_thresh); in CollectHistogram_SSE41()
52 _mm_storeu_si128((__m128i*)&out[0], bin0); in CollectHistogram_SSE41()
[all …]
Ddec_sse2.c50 const __m128i k1 = _mm_set1_epi16(20091); in Transform_SSE2()
51 const __m128i k2 = _mm_set1_epi16(-30068); in Transform_SSE2()
52 __m128i T0, T1, T2, T3; in Transform_SSE2()
57 __m128i in0, in1, in2, in3; in Transform_SSE2()
59 in0 = _mm_loadl_epi64((const __m128i*)&in[0]); in Transform_SSE2()
60 in1 = _mm_loadl_epi64((const __m128i*)&in[4]); in Transform_SSE2()
61 in2 = _mm_loadl_epi64((const __m128i*)&in[8]); in Transform_SSE2()
62 in3 = _mm_loadl_epi64((const __m128i*)&in[12]); in Transform_SSE2()
68 const __m128i inB0 = _mm_loadl_epi64((const __m128i*)&in[16]); in Transform_SSE2()
69 const __m128i inB1 = _mm_loadl_epi64((const __m128i*)&in[20]); in Transform_SSE2()
[all …]
Dyuv_sse2.c29 static void ConvertYUV444ToRGB_SSE2(const __m128i* const Y0, in ConvertYUV444ToRGB_SSE2()
30 const __m128i* const U0, in ConvertYUV444ToRGB_SSE2()
31 const __m128i* const V0, in ConvertYUV444ToRGB_SSE2()
32 __m128i* const R, in ConvertYUV444ToRGB_SSE2()
33 __m128i* const G, in ConvertYUV444ToRGB_SSE2()
34 __m128i* const B) { in ConvertYUV444ToRGB_SSE2()
35 const __m128i k19077 = _mm_set1_epi16(19077); in ConvertYUV444ToRGB_SSE2()
36 const __m128i k26149 = _mm_set1_epi16(26149); in ConvertYUV444ToRGB_SSE2()
37 const __m128i k14234 = _mm_set1_epi16(14234); in ConvertYUV444ToRGB_SSE2()
39 const __m128i k33050 = _mm_set1_epi16((short)33050); in ConvertYUV444ToRGB_SSE2()
[all …]
Drescaler_sse2.c31 static void LoadTwoPixels_SSE2(const uint8_t* const src, __m128i* out) { in LoadTwoPixels_SSE2()
32 const __m128i zero = _mm_setzero_si128(); in LoadTwoPixels_SSE2()
33 const __m128i A = _mm_loadl_epi64((const __m128i*)(src)); // ABCDEFGH in LoadTwoPixels_SSE2()
34 const __m128i B = _mm_unpacklo_epi8(A, zero); // A0B0C0D0E0F0G0H0 in LoadTwoPixels_SSE2()
35 const __m128i C = _mm_srli_si128(B, 8); // E0F0G0H0 in LoadTwoPixels_SSE2()
40 static void LoadEightPixels_SSE2(const uint8_t* const src, __m128i* out) { in LoadEightPixels_SSE2()
41 const __m128i zero = _mm_setzero_si128(); in LoadEightPixels_SSE2()
42 const __m128i A = _mm_loadl_epi64((const __m128i*)(src)); // ABCDEFGH in LoadEightPixels_SSE2()
52 __m128i cur_pixels; in RescalerImportRowExpand_SSE2()
66 const __m128i mult = _mm_set1_epi32(((x_add - accum) << 16) | accum); in RescalerImportRowExpand_SSE2()
[all …]
Dyuv_sse41.c29 static void ConvertYUV444ToRGB_SSE41(const __m128i* const Y0, in ConvertYUV444ToRGB_SSE41()
30 const __m128i* const U0, in ConvertYUV444ToRGB_SSE41()
31 const __m128i* const V0, in ConvertYUV444ToRGB_SSE41()
32 __m128i* const R, in ConvertYUV444ToRGB_SSE41()
33 __m128i* const G, in ConvertYUV444ToRGB_SSE41()
34 __m128i* const B) { in ConvertYUV444ToRGB_SSE41()
35 const __m128i k19077 = _mm_set1_epi16(19077); in ConvertYUV444ToRGB_SSE41()
36 const __m128i k26149 = _mm_set1_epi16(26149); in ConvertYUV444ToRGB_SSE41()
37 const __m128i k14234 = _mm_set1_epi16(14234); in ConvertYUV444ToRGB_SSE41()
39 const __m128i k33050 = _mm_set1_epi16((short)33050); in ConvertYUV444ToRGB_SSE41()
[all …]
Dcommon_sse2.h30 static WEBP_INLINE void PrintReg(const __m128i r, const char* const name,
34 __m128i r;
59 static WEBP_INLINE int VP8HorizontalAdd8b(const __m128i* const a) { in VP8HorizontalAdd8b()
60 const __m128i zero = _mm_setzero_si128(); in VP8HorizontalAdd8b()
61 const __m128i sad8x2 = _mm_sad_epu8(*a, zero); in VP8HorizontalAdd8b()
63 const __m128i sum = _mm_add_epi32(sad8x2, _mm_shuffle_epi32(sad8x2, 2)); in VP8HorizontalAdd8b()
69 const __m128i* const in0, const __m128i* const in1, in VP8Transpose_2_4x4_16b()
70 const __m128i* const in2, const __m128i* const in3, __m128i* const out0, in VP8Transpose_2_4x4_16b()
71 __m128i* const out1, __m128i* const out2, __m128i* const out3) { in VP8Transpose_2_4x4_16b()
77 const __m128i transpose0_0 = _mm_unpacklo_epi16(*in0, *in1); in VP8Transpose_2_4x4_16b()
[all …]
Dlossless_enc_sse41.c30 const __m128i kCstShuffle = _mm_set_epi8(-1, 13, -1, 13, -1, 9, -1, 9, in SubtractGreenFromBlueAndRed_SSE41()
33 const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); in SubtractGreenFromBlueAndRed_SSE41()
34 const __m128i in_0g0g = _mm_shuffle_epi8(in, kCstShuffle); in SubtractGreenFromBlueAndRed_SSE41()
35 const __m128i out = _mm_sub_epi8(in, in_0g0g); in SubtractGreenFromBlueAndRed_SSE41()
36 _mm_storeu_si128((__m128i*)&argb_data[i], out); in SubtractGreenFromBlueAndRed_SSE41()
52 const __m128i mults_r = _mm_set1_epi16(CST_5b(red_to_blue)); in CollectColorBlueTransforms_SSE41()
53 const __m128i mults_g = _mm_set1_epi16(CST_5b(green_to_blue)); in CollectColorBlueTransforms_SSE41()
54 const __m128i mask_g = _mm_set1_epi16((short)0xff00); // green mask in CollectColorBlueTransforms_SSE41()
55 const __m128i mask_gb = _mm_set1_epi32(0xffff); // green/blue mask in CollectColorBlueTransforms_SSE41()
56 const __m128i mask_b = _mm_set1_epi16(0x00ff); // blue mask in CollectColorBlueTransforms_SSE41()
[all …]
/third_party/mesa3d/src/gallium/auxiliary/util/
Du_pwr8.h39 typedef VECTOR_ALIGN_16 vector unsigned char __m128i; typedef
42 __m128i m128i;
51 static inline __m128i
68 return (__m128i) vdst.m128si; in vec_set_epi32()
71 static inline __m128i
77 static inline __m128i
78 vec_unpacklo_epi32 (__m128i even, __m128i odd) in vec_unpacklo_epi32()
80 static const __m128i perm_mask = in vec_unpacklo_epi32()
90 static inline __m128i
91 vec_unpackhi_epi32 (__m128i even, __m128i odd) in vec_unpackhi_epi32()
[all …]
Du_sse.h50 __m128i m;
56 static inline void u_print_epi8(const char *name, __m128i r) in u_print_epi8()
58 union { __m128i m; ubyte ub[16]; } u; in u_print_epi8()
85 static inline void u_print_epi16(const char *name, __m128i r) in u_print_epi16()
87 union { __m128i m; ushort us[8]; } u; in u_print_epi16()
104 static inline void u_print_epi32(const char *name, __m128i r) in u_print_epi32()
106 union { __m128i m; uint ui[4]; } u; in u_print_epi32()
153 static __inline __m128i
159 _mm_shuffle_epi8(__m128i a, __m128i mask) in _mm_shuffle_epi8()
161 __m128i result; in _mm_shuffle_epi8()
[all …]
/third_party/mesa3d/src/gallium/drivers/llvmpipe/
Dlp_rast_tri.c177 __m128i cstep0 = _mm_setr_epi32(c, c+dcdx, c+dcdx*2, c+dcdx*3); in build_masks_sse()
178 __m128i xdcdy = _mm_set1_epi32(dcdy); in build_masks_sse()
182 __m128i cstep1 = _mm_add_epi32(cstep0, xdcdy); in build_masks_sse()
183 __m128i cstep2 = _mm_add_epi32(cstep1, xdcdy); in build_masks_sse()
184 __m128i cstep3 = _mm_add_epi32(cstep2, xdcdy); in build_masks_sse()
187 __m128i cstep01, cstep23, result; in build_masks_sse()
198 __m128i cio4 = _mm_set1_epi32(cdiff); in build_masks_sse()
199 __m128i cstep01, cstep23, result; in build_masks_sse()
218 __m128i cstep0 = _mm_setr_epi32(c, c+dcdx, c+dcdx*2, c+dcdx*3); in build_mask_linear_sse()
219 __m128i xdcdy = _mm_set1_epi32(dcdy); in build_mask_linear_sse()
[all …]

1234567