PR #25670 opened by Kacper Michajłow (kasper93) URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/25670 Patch URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/25670.patch
From 6b2c1a4166e0d222d5e062b85108b243dc87a5e8 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kacper=20Michaj=C5=82ow?= <[email protected]> Date: Thu, 29 Oct 2026 10:44:22 +0100 Subject: [PATCH 1/2] swscale/uops: round every float operation of the reference backend MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit With x87 math the compiler keeps the intermediate results of an expression with excess precision. Round them, so that they produce the same results. Signed-off-by: Kacper Michajłow <[email protected]> --- libswscale/uops.h | 2 +- libswscale/uops_tmpl.c | 52 +++++++++++++++++++++++++++--------------- 2 files changed, 34 insertions(+), 20 deletions(-) diff --git a/libswscale/uops.h b/libswscale/uops.h index 506b3ff6222..4d25107927b 100644 --- a/libswscale/uops.h +++ b/libswscale/uops.h @@ -94,7 +94,7 @@ static inline SwsPixel ff_sws_pixel_from_q64(SwsPixelType type, AVRational64 val case SWS_PIXEL_U8: return (SwsPixel) { .u8 = val.num / val.den }; case SWS_PIXEL_U16: return (SwsPixel) { .u16 = val.num / val.den }; case SWS_PIXEL_U32: return (SwsPixel) { .u32 = val.num / val.den }; - case SWS_PIXEL_F32: return (SwsPixel) { .f32 = (float) val.num / val.den }; + case SWS_PIXEL_F32: return (SwsPixel) { .f32 = (float) val.num / (float) val.den }; case SWS_PIXEL_NONE: case SWS_PIXEL_TYPE_NB: break; } diff --git a/libswscale/uops_tmpl.c b/libswscale/uops_tmpl.c index e94bfca417f..56e28749b12 100644 --- a/libswscale/uops_tmpl.c +++ b/libswscale/uops_tmpl.c @@ -61,6 +61,9 @@ # error Invalid BIT_DEPTH #endif +/* Every intermediate float result is cast to pixel_t, which rounds away the + * excess precision of x87 math for the same results as everywhere else */ + /********************************* * Generic read/write operations * *********************************/ @@ -285,10 +288,10 @@ DECL_READ(read_planar_fv, const SwsCompMask mask, const SwsPixelType type) SWS_LOOP for (int i = 0; i < SWS_BLOCK_SIZE; i++) { - if (X) xs.f32[i] += weight * in0[i]; - if (Y) ys.f32[i] += weight * in1[i]; - if (Z) zs.f32[i] += weight * in2[i]; - if (W) ws.f32[i] += weight * in3[i]; + if (X) xs.f32[i] += (float) (weight * in0[i]); + if (Y) ys.f32[i] += (float) (weight * in1[i]); + if (Z) zs.f32[i] += (float) (weight * in2[i]); + if (W) ws.f32[i] += (float) (weight * in3[i]); } if (X) in0 = bump_ptr(in0, exec->in_stride[0]); @@ -355,10 +358,10 @@ DECL_READ(read_planar_fh, const SwsCompMask mask, const SwsPixelType type) inter_t sx = 0, sy = 0, sz = 0, sw = 0; for (int j = 0; j < filter_size; j++) { const int weight = weights[j]; - if (X) sx += weight * start0[j]; - if (Y) sy += weight * start1[j]; - if (Z) sz += weight * start2[j]; - if (W) sw += weight * start3[j]; + if (X) sx += (inter_t) (weight * start0[j]); + if (Y) sy += (inter_t) (weight * start1[j]); + if (Z) sz += (inter_t) (weight * start2[j]); + if (W) sw += (inter_t) (weight * start3[j]); } if (X) xs.f32[i] = (float) sx * scale; @@ -838,7 +841,7 @@ DECL_FUNC(linear, const SwsCompMask mask, const uint32_t one, const uint32_t zer const pixel_t ww = w[i]; #define LIN_VAL(I, J, val) \ - ((one & SWS_MASK(I, J)) ? (val) : (uinter_t) c.m[I][J] * (val)) + ((one & SWS_MASK(I, J)) ? (val) : (pixel_t) ((uinter_t) c.m[I][J] * (val))) #define LIN_ROW(I, var) do { \ pixel_t tmp = (zero & SWS_MASK(I, 4)) ? 0 : c.k[I]; \ @@ -879,7 +882,18 @@ av_always_inline static vec3_t fn(vec3)(v3u16_t v) return (vec3_t) { v.x, v.y, v.z }; } -#define lerp(a, b, w) ((a) + (w) * ((pixel_t) (b) - (a))) +#define lerp(a, b, w) ((pixel_t) ((a) + (pixel_t) ((w) * (pixel_t) ((pixel_t) (b) - (a))))) + +av_always_inline static +pixel_t fn(sum4)(pixel_t w0, pixel_t a, pixel_t w1, pixel_t b, + pixel_t w2, pixel_t c, pixel_t w3, pixel_t d) +{ + pixel_t sum = (pixel_t) (w0 * a); + sum += (pixel_t) (w1 * b); + sum += (pixel_t) (w2 * c); + sum += (pixel_t) (w3 * d); + return sum; +} av_always_inline static vec3_t fn(lerp3)(vec3_t a, vec3_t b, pixel_t w) @@ -932,16 +946,16 @@ vec3_t fn(lut3d_static)(const SwsLut3D *restrict lut3d, vec3_t rgb) const vec3_t v3 = fn(vec3)(base[off0 + off1 + off2]); return (vec3_t) { - w0 * v0.x + w1 * v1.x + w2 * v2.x + w3 * v3.x, - w0 * v0.y + w1 * v1.y + w2 * v2.y + w3 * v3.y, - w0 * v0.z + w1 * v1.z + w2 * v2.z + w3 * v3.z, + fn(sum4)(w0, v0.x, w1, v1.x, w2, v2.x, w3, v3.x), + fn(sum4)(w0, v0.y, w1, v1.y, w2, v2.y, w3, v3.y), + fn(sum4)(w0, v0.z, w1, v1.z, w2, v2.z, w3, v3.z), }; } av_always_inline static vec3_t fn(lut3d_dynamic)(const SwsLut3D *restrict lut3d, vec3_t rgb) { - rgb.x *= (TONE_LUT_SIZE - 1) / (pixel_t) UINT16_MAX; + rgb.x *= (pixel_t) ((TONE_LUT_SIZE - 1) / (pixel_t) UINT16_MAX); /* Linear interpolation */ const int Ix = (int) rgb.x; @@ -955,13 +969,13 @@ vec3_t fn(lut3d_dynamic)(const SwsLut3D *restrict lut3d, vec3_t rgb) const pixel_t scale = k / (pixel_t) (1 << 15); rgb.x = lerp(a.x, b.x, If); - rgb.y = bias + scale * rgb.y; - rgb.z = bias + scale * rgb.z; + rgb.y = bias + (pixel_t) (scale * rgb.y); + rgb.z = bias + (pixel_t) (scale * rgb.z); /* Re-scale to output LUT size */ - rgb.x *= (OUTPUT_LUT_SIZE_I - 1) / (pixel_t) UINT16_MAX; - rgb.y *= (OUTPUT_LUT_SIZE_PT - 1) / (pixel_t) UINT16_MAX; - rgb.z *= (OUTPUT_LUT_SIZE_PT - 1) / (pixel_t) UINT16_MAX; + rgb.x *= (pixel_t) ((OUTPUT_LUT_SIZE_I - 1) / (pixel_t) UINT16_MAX); + rgb.y *= (pixel_t) ((OUTPUT_LUT_SIZE_PT - 1) / (pixel_t) UINT16_MAX); + rgb.z *= (pixel_t) ((OUTPUT_LUT_SIZE_PT - 1) / (pixel_t) UINT16_MAX); /* Trilinear interpolation */ const int lo0 = (int) rgb.x; -- 2.54.0 From 03ef9aba552e65cc336d2a53c690d10d5cbe0d94 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kacper=20Michaj=C5=82ow?= <[email protected]> Date: Thu, 29 Oct 2026 18:38:17 +0100 Subject: [PATCH 2/2] swscale/uops: check every product of an input before fusing it MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit exact_prod() only checked the products with the two ends of the input range, so a coefficient times inputs in [0, 1] or [0, 2] passed as exact although the values in between are not, and with SWS_BITEXACT the x86 backend fused those products with FMA where the C backend rounds them. A product is exact for every input when the coefficient is a power of two, or when the inputs are integers whose bits fit in the significand next to those of the coefficient. Signed-off-by: Kacper Michajłow <[email protected]> --- libswscale/uops.c | 30 +++++++++++++++++++----------- libswscale/x86/uops_macros.asm | 16 ++++++++++++++-- libswscale/x86/uops_macros.h | 32 ++++++++++++++++++++++++++++---- 3 files changed, 61 insertions(+), 17 deletions(-) diff --git a/libswscale/uops.c b/libswscale/uops.c index a7150c11307..5fcfd6dea21 100644 --- a/libswscale/uops.c +++ b/libswscale/uops.c @@ -18,9 +18,12 @@ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA */ +#include <float.h> +#include <math.h> #include <stdbool.h> #include "libavutil/avassert.h" +#include "libavutil/intmath.h" #include "libavutil/mem.h" #include "libavutil/refstruct.h" @@ -236,26 +239,31 @@ static SwsPixelType pixel_type_to_int(SwsUOpFlags flags, const SwsPixelType type return SWS_PIXEL_NONE; } -static bool exact_product_f32(float a, float b) +static int significant_bits_f32(float f) { - volatile float prod = a * b; - volatile float result = b ? prod / b : 0.0f; - return !b || result == a; + int exp; + const uint32_t mant = ldexpf(frexpf(fabsf(f), &exp), FLT_MANT_DIG); + return mant ? FLT_MANT_DIG - ff_ctz(mant) : 0; } static bool exact_prod(SwsPixel coef, const SwsOpPad *input) { if (ff_sws_pixel_type_is_int(input->type)) return true; - else if (!input->min.den || !input->max.den) - return false; /* unknown bounds */ - const SwsPixel min = ff_sws_pixel_from_q64(input->type, input->min); - const SwsPixel max = ff_sws_pixel_from_q64(input->type, input->max); switch (input->type) { - case SWS_PIXEL_F32: - return exact_product_f32(coef.f32, min.f32) && - exact_product_f32(coef.f32, max.f32); + case SWS_PIXEL_F32: { + const int bits = significant_bits_f32(coef.f32); + if (bits <= 1) + return true; /* power of two */ + if (!(input->flags & SWS_COMP_EXACT) || !input->min.den || !input->max.den) + return false; /* not only integers, or unknown bounds */ + + const double max = FFMAX(fabs(ff_q2d_64(input->min)), fabs(ff_q2d_64(input->max))); + if (max >= 1 << FLT_MANT_DIG) + return false; + return bits + (max >= 1.0 ? av_log2((unsigned) max) + 1 : 0) <= FLT_MANT_DIG; + } } av_unreachable("Invalid pixel type!"); diff --git a/libswscale/x86/uops_macros.asm b/libswscale/x86/uops_macros.asm index 09bc564ba3e..ba002f0edcf 100644 --- a/libswscale/x86/uops_macros.asm +++ b/libswscale/x86/uops_macros.asm @@ -825,6 +825,7 @@ %macro DECL_F32_LINEAR_FMA 1 DECL_OP %1, f32_linear_fma_x_xxx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x1, 0x41040, 0xbefa8, 0xfffe8 + DECL_OP %1, f32_linear_fma_x_xXx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x1, 0x41040, 0xbefa8, 0xfffea DECL_OP %1, f32_linear_fma_x_XXX0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x1, 0x41040, 0xbefa8, 0xfffef DECL_OP %1, f32_linear_fma_x_x000x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x1, 0x41040, 0xbefae, 0xfffee DECL_OP %1, f32_linear_fma_x_X000x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x1, 0x41040, 0xbefae, 0xfffef @@ -834,14 +835,24 @@ DECL_OP %1, f32_linear_fma_x_xXX00 , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x1, 0x41040, 0xbefb8, 0xffffe DECL_OP %1, f32_linear_fma_y_0x000 , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x2, 0x41001, 0xbefbe, 0xfffbf DECL_OP %1, f32_linear_fma_xyz_xxx0x_xxx0x_xxx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba108, 0xfa108 - DECL_OP %1, f32_linear_fma_xyz_XXX0x_XxX0x_XXX0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba108, 0xfbdaf + DECL_OP %1, f32_linear_fma_xyz_xxx0x_xxx0x_Xxx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba108, 0xfa508 + DECL_OP %1, f32_linear_fma_xyz_xxx0x_xxx0x_XXx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba108, 0xfad08 + DECL_OP %1, f32_linear_fma_xyz_XxX0x_xxX0x_Xxx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba108, 0xfa58d DECL_OP %1, f32_linear_fma_xyz_XXX0x_XXX0x_XXX0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba108, 0xfbdef + DECL_OP %1, f32_linear_fma_xyz_x0x0x_xxx0x_xx00x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xbb10a, 0xfb10a + DECL_OP %1, f32_linear_fma_xyz_X0x0x_Xxx0x_Xx00x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xbb10a, 0xfb52b DECL_OP %1, f32_linear_fma_xyz_X0X0x_XXX0x_XX00x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xbb10a, 0xfbdef DECL_OP %1, f32_linear_fma_xyz_xxx00_xxx0x_xxx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba118, 0xfa118 - DECL_OP %1, f32_linear_fma_xyz_xXX00_XxX0x_XXX0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba118, 0xfbdbe + DECL_OP %1, f32_linear_fma_xyz_xxx00_xXx0x_xxx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba118, 0xfa158 + DECL_OP %1, f32_linear_fma_xyz_xxx00_xxX0x_xxx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba118, 0xfa198 + DECL_OP %1, f32_linear_fma_xyz_xxx00_xxX0x_Xxx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba118, 0xfa598 + DECL_OP %1, f32_linear_fma_xyz_xxX00_xxX0x_XxX0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba118, 0xfb59c DECL_OP %1, f32_linear_fma_xyz_xXX00_XXX0x_XXX0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba118, 0xfbdfe + DECL_OP %1, f32_linear_fma_xyz_x000x_0x00x_00x0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xbadae, 0xfadae + DECL_OP %1, f32_linear_fma_xyz_X000x_0x00x_00X0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xbadae, 0xfbdaf DECL_OP %1, f32_linear_fma_xyz_X000x_0X00x_00X0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xbadae, 0xfbdef DECL_OP %1, f32_linear_fma_xyz_x0000_0x000_00x00 , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xbefbe, 0xfefbe + DECL_OP %1, f32_linear_fma_xyz_10x0x_1xx0x_1x00x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40421, 0xbb10a, 0xfb52b DECL_OP %1, f32_linear_fma_xyz_10X0x_1XX0x_1X00x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40421, 0xbb10a, 0xfbdef DECL_OP %1, f32_linear_fma_w_000x0 , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x8, 0x01041, 0xbefbe, 0xbffff DECL_OP %1, f32_linear_fma_xw_x000x_000x0 , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x9, 0x01040, 0xbefae, 0xbffee @@ -850,6 +861,7 @@ DECL_OP %1, f32_linear_fma_xw_xXX00_000x0 , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x9, 0x01040, 0xbefb8, 0xbfffe DECL_OP %1, f32_linear_fma_xyzw_xxx0x_xxx0x_xxx0x_000x0, SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0xf, 0x00000, 0xba108, 0xba108 DECL_OP %1, f32_linear_fma_xyzw_XXX0x_XXX0x_XXX0x_000x0, SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0xf, 0x00000, 0xba108, 0xbbdef + DECL_OP %1, f32_linear_fma_xyzw_x0x0x_xxx0x_xx00x_000x0, SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0xf, 0x00000, 0xbb10a, 0xbb10a DECL_OP %1, f32_linear_fma_xyzw_X0X0x_XXX0x_XX00x_000x0, SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0xf, 0x00000, 0xbb10a, 0xbbdef DECL_OP %1, f32_linear_fma_xyzw_x0000_0x000_00x00_000x0, SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0xf, 0x00000, 0xbefbe, 0xbefbe %endmacro diff --git a/libswscale/x86/uops_macros.h b/libswscale/x86/uops_macros.h index 66b3a6afe27..b3312873713 100644 --- a/libswscale/x86/uops_macros.h +++ b/libswscale/x86/uops_macros.h @@ -1087,6 +1087,7 @@ MACRO(__VA_ARGS__, f32_linear_xyzw_x0000_0x000_00x00_000x0 , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR , .mask = 0xf, .par.lin.one = 0x00000, .par.lin.zero = 0xbefbe) #define SWS_FOR_F32_LINEAR_FMA(MACRO, ...) \ MACRO(__VA_ARGS__, f32_linear_fma_x_xxx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x1, 0x41040, 0xbefa8, 0xfffe8) \ + MACRO(__VA_ARGS__, f32_linear_fma_x_xXx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x1, 0x41040, 0xbefa8, 0xfffea) \ MACRO(__VA_ARGS__, f32_linear_fma_x_XXX0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x1, 0x41040, 0xbefa8, 0xfffef) \ MACRO(__VA_ARGS__, f32_linear_fma_x_x000x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x1, 0x41040, 0xbefae, 0xfffee) \ MACRO(__VA_ARGS__, f32_linear_fma_x_X000x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x1, 0x41040, 0xbefae, 0xfffef) \ @@ -1096,14 +1097,24 @@ MACRO(__VA_ARGS__, f32_linear_fma_x_xXX00 , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x1, 0x41040, 0xbefb8, 0xffffe) \ MACRO(__VA_ARGS__, f32_linear_fma_y_0x000 , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x2, 0x41001, 0xbefbe, 0xfffbf) \ MACRO(__VA_ARGS__, f32_linear_fma_xyz_xxx0x_xxx0x_xxx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba108, 0xfa108) \ - MACRO(__VA_ARGS__, f32_linear_fma_xyz_XXX0x_XxX0x_XXX0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba108, 0xfbdaf) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_xxx0x_xxx0x_Xxx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba108, 0xfa508) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_xxx0x_xxx0x_XXx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba108, 0xfad08) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_XxX0x_xxX0x_Xxx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba108, 0xfa58d) \ MACRO(__VA_ARGS__, f32_linear_fma_xyz_XXX0x_XXX0x_XXX0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba108, 0xfbdef) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_x0x0x_xxx0x_xx00x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xbb10a, 0xfb10a) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_X0x0x_Xxx0x_Xx00x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xbb10a, 0xfb52b) \ MACRO(__VA_ARGS__, f32_linear_fma_xyz_X0X0x_XXX0x_XX00x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xbb10a, 0xfbdef) \ MACRO(__VA_ARGS__, f32_linear_fma_xyz_xxx00_xxx0x_xxx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba118, 0xfa118) \ - MACRO(__VA_ARGS__, f32_linear_fma_xyz_xXX00_XxX0x_XXX0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba118, 0xfbdbe) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_xxx00_xXx0x_xxx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba118, 0xfa158) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_xxx00_xxX0x_xxx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba118, 0xfa198) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_xxx00_xxX0x_Xxx0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba118, 0xfa598) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_xxX00_xxX0x_XxX0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba118, 0xfb59c) \ MACRO(__VA_ARGS__, f32_linear_fma_xyz_xXX00_XXX0x_XXX0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xba118, 0xfbdfe) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_x000x_0x00x_00x0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xbadae, 0xfadae) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_X000x_0x00x_00X0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xbadae, 0xfbdaf) \ MACRO(__VA_ARGS__, f32_linear_fma_xyz_X000x_0X00x_00X0x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xbadae, 0xfbdef) \ MACRO(__VA_ARGS__, f32_linear_fma_xyz_x0000_0x000_00x00 , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40000, 0xbefbe, 0xfefbe) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_10x0x_1xx0x_1x00x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40421, 0xbb10a, 0xfb52b) \ MACRO(__VA_ARGS__, f32_linear_fma_xyz_10X0x_1XX0x_1X00x , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x7, 0x40421, 0xbb10a, 0xfbdef) \ MACRO(__VA_ARGS__, f32_linear_fma_w_000x0 , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x8, 0x01041, 0xbefbe, 0xbffff) \ MACRO(__VA_ARGS__, f32_linear_fma_xw_x000x_000x0 , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x9, 0x01040, 0xbefae, 0xbffee) \ @@ -1112,10 +1123,12 @@ MACRO(__VA_ARGS__, f32_linear_fma_xw_xXX00_000x0 , SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0x9, 0x01040, 0xbefb8, 0xbfffe) \ MACRO(__VA_ARGS__, f32_linear_fma_xyzw_xxx0x_xxx0x_xxx0x_000x0, SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0xf, 0x00000, 0xba108, 0xba108) \ MACRO(__VA_ARGS__, f32_linear_fma_xyzw_XXX0x_XXX0x_XXX0x_000x0, SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0xf, 0x00000, 0xba108, 0xbbdef) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyzw_x0x0x_xxx0x_xx00x_000x0, SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0xf, 0x00000, 0xbb10a, 0xbb10a) \ MACRO(__VA_ARGS__, f32_linear_fma_xyzw_X0X0x_XXX0x_XX00x_000x0, SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0xf, 0x00000, 0xbb10a, 0xbbdef) \ MACRO(__VA_ARGS__, f32_linear_fma_xyzw_x0000_0x000_00x00_000x0, SWS_PIXEL_F32, SWS_UOP_LINEAR_FMA , 0xf, 0x00000, 0xbefbe, 0xbefbe) #define SWS_FOR_STRUCT_F32_LINEAR_FMA(MACRO, ...) \ MACRO(__VA_ARGS__, f32_linear_fma_x_xxx0x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x1, .par.lin.one = 0x41040, .par.lin.zero = 0xbefa8, .par.lin.exact = 0xfffe8) \ + MACRO(__VA_ARGS__, f32_linear_fma_x_xXx0x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x1, .par.lin.one = 0x41040, .par.lin.zero = 0xbefa8, .par.lin.exact = 0xfffea) \ MACRO(__VA_ARGS__, f32_linear_fma_x_XXX0x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x1, .par.lin.one = 0x41040, .par.lin.zero = 0xbefa8, .par.lin.exact = 0xfffef) \ MACRO(__VA_ARGS__, f32_linear_fma_x_x000x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x1, .par.lin.one = 0x41040, .par.lin.zero = 0xbefae, .par.lin.exact = 0xfffee) \ MACRO(__VA_ARGS__, f32_linear_fma_x_X000x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x1, .par.lin.one = 0x41040, .par.lin.zero = 0xbefae, .par.lin.exact = 0xfffef) \ @@ -1125,14 +1138,24 @@ MACRO(__VA_ARGS__, f32_linear_fma_x_xXX00 , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x1, .par.lin.one = 0x41040, .par.lin.zero = 0xbefb8, .par.lin.exact = 0xffffe) \ MACRO(__VA_ARGS__, f32_linear_fma_y_0x000 , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x2, .par.lin.one = 0x41001, .par.lin.zero = 0xbefbe, .par.lin.exact = 0xfffbf) \ MACRO(__VA_ARGS__, f32_linear_fma_xyz_xxx0x_xxx0x_xxx0x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xba108, .par.lin.exact = 0xfa108) \ - MACRO(__VA_ARGS__, f32_linear_fma_xyz_XXX0x_XxX0x_XXX0x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xba108, .par.lin.exact = 0xfbdaf) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_xxx0x_xxx0x_Xxx0x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xba108, .par.lin.exact = 0xfa508) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_xxx0x_xxx0x_XXx0x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xba108, .par.lin.exact = 0xfad08) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_XxX0x_xxX0x_Xxx0x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xba108, .par.lin.exact = 0xfa58d) \ MACRO(__VA_ARGS__, f32_linear_fma_xyz_XXX0x_XXX0x_XXX0x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xba108, .par.lin.exact = 0xfbdef) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_x0x0x_xxx0x_xx00x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xbb10a, .par.lin.exact = 0xfb10a) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_X0x0x_Xxx0x_Xx00x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xbb10a, .par.lin.exact = 0xfb52b) \ MACRO(__VA_ARGS__, f32_linear_fma_xyz_X0X0x_XXX0x_XX00x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xbb10a, .par.lin.exact = 0xfbdef) \ MACRO(__VA_ARGS__, f32_linear_fma_xyz_xxx00_xxx0x_xxx0x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xba118, .par.lin.exact = 0xfa118) \ - MACRO(__VA_ARGS__, f32_linear_fma_xyz_xXX00_XxX0x_XXX0x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xba118, .par.lin.exact = 0xfbdbe) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_xxx00_xXx0x_xxx0x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xba118, .par.lin.exact = 0xfa158) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_xxx00_xxX0x_xxx0x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xba118, .par.lin.exact = 0xfa198) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_xxx00_xxX0x_Xxx0x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xba118, .par.lin.exact = 0xfa598) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_xxX00_xxX0x_XxX0x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xba118, .par.lin.exact = 0xfb59c) \ MACRO(__VA_ARGS__, f32_linear_fma_xyz_xXX00_XXX0x_XXX0x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xba118, .par.lin.exact = 0xfbdfe) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_x000x_0x00x_00x0x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xbadae, .par.lin.exact = 0xfadae) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_X000x_0x00x_00X0x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xbadae, .par.lin.exact = 0xfbdaf) \ MACRO(__VA_ARGS__, f32_linear_fma_xyz_X000x_0X00x_00X0x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xbadae, .par.lin.exact = 0xfbdef) \ MACRO(__VA_ARGS__, f32_linear_fma_xyz_x0000_0x000_00x00 , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40000, .par.lin.zero = 0xbefbe, .par.lin.exact = 0xfefbe) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyz_10x0x_1xx0x_1x00x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40421, .par.lin.zero = 0xbb10a, .par.lin.exact = 0xfb52b) \ MACRO(__VA_ARGS__, f32_linear_fma_xyz_10X0x_1XX0x_1X00x , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x7, .par.lin.one = 0x40421, .par.lin.zero = 0xbb10a, .par.lin.exact = 0xfbdef) \ MACRO(__VA_ARGS__, f32_linear_fma_w_000x0 , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x8, .par.lin.one = 0x01041, .par.lin.zero = 0xbefbe, .par.lin.exact = 0xbffff) \ MACRO(__VA_ARGS__, f32_linear_fma_xw_x000x_000x0 , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x9, .par.lin.one = 0x01040, .par.lin.zero = 0xbefae, .par.lin.exact = 0xbffee) \ @@ -1141,6 +1164,7 @@ MACRO(__VA_ARGS__, f32_linear_fma_xw_xXX00_000x0 , .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0x9, .par.lin.one = 0x01040, .par.lin.zero = 0xbefb8, .par.lin.exact = 0xbfffe) \ MACRO(__VA_ARGS__, f32_linear_fma_xyzw_xxx0x_xxx0x_xxx0x_000x0, .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0xf, .par.lin.one = 0x00000, .par.lin.zero = 0xba108, .par.lin.exact = 0xba108) \ MACRO(__VA_ARGS__, f32_linear_fma_xyzw_XXX0x_XXX0x_XXX0x_000x0, .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0xf, .par.lin.one = 0x00000, .par.lin.zero = 0xba108, .par.lin.exact = 0xbbdef) \ + MACRO(__VA_ARGS__, f32_linear_fma_xyzw_x0x0x_xxx0x_xx00x_000x0, .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0xf, .par.lin.one = 0x00000, .par.lin.zero = 0xbb10a, .par.lin.exact = 0xbb10a) \ MACRO(__VA_ARGS__, f32_linear_fma_xyzw_X0X0x_XXX0x_XX00x_000x0, .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0xf, .par.lin.one = 0x00000, .par.lin.zero = 0xbb10a, .par.lin.exact = 0xbbdef) \ MACRO(__VA_ARGS__, f32_linear_fma_xyzw_x0000_0x000_00x00_000x0, .type = SWS_PIXEL_F32, .uop = SWS_UOP_LINEAR_FMA , .mask = 0xf, .par.lin.one = 0x00000, .par.lin.zero = 0xbefbe, .par.lin.exact = 0xbefbe) #define SWS_FOR_F32_DITHER(MACRO, ...) \ -- 2.54.0 _______________________________________________ ffmpeg-devel mailing list -- [email protected] To unsubscribe send an email to [email protected]
