swscale/x86/input.asm: add x86-optimized planer rgb2yuv functions

sse2 only operates on 2 lanes per loop for to_y and to_uv functions, due
to the lack of pmulld instruction.  Emulating pmulld with 2 pmuludq and shuffles
proved too costly and made to_uv functions slower then the c implementation.

For to_y on sse2 only float functions are generated,
I was are not able outperform the c implementation on the integer pixel formats.

For to_a on see4 only the float functions are generated.
sse2 and sse4 generated nearly identical performing code on integer pixel formats,
so only sse2/avx2 versions are generated.

planar_gbrp_to_y_512_c: 1197.5
planar_gbrp_to_y_512_sse4: 444.5
planar_gbrp_to_y_512_avx2: 287.5
planar_gbrap_to_y_512_c: 1204.5
planar_gbrap_to_y_512_sse4: 447.5
planar_gbrap_to_y_512_avx2: 289.5
planar_gbrp9be_to_y_512_c: 1380.0
planar_gbrp9be_to_y_512_sse4: 543.5
planar_gbrp9be_to_y_512_avx2: 340.0
planar_gbrp9le_to_y_512_c: 1200.5
planar_gbrp9le_to_y_512_sse4: 442.0
planar_gbrp9le_to_y_512_avx2: 282.0
planar_gbrp10be_to_y_512_c: 1378.5
planar_gbrp10be_to_y_512_sse4: 544.0
planar_gbrp10be_to_y_512_avx2: 337.5
planar_gbrp10le_to_y_512_c: 1200.0
planar_gbrp10le_to_y_512_sse4: 448.0
planar_gbrp10le_to_y_512_avx2: 285.5
planar_gbrap10be_to_y_512_c: 1380.0
planar_gbrap10be_to_y_512_sse4: 542.0
planar_gbrap10be_to_y_512_avx2: 340.5
planar_gbrap10le_to_y_512_c: 1199.0
planar_gbrap10le_to_y_512_sse4: 446.0
planar_gbrap10le_to_y_512_avx2: 289.5
planar_gbrp12be_to_y_512_c: 10563.0
planar_gbrp12be_to_y_512_sse4: 542.5
planar_gbrp12be_to_y_512_avx2: 339.0
planar_gbrp12le_to_y_512_c: 1201.0
planar_gbrp12le_to_y_512_sse4: 440.5
planar_gbrp12le_to_y_512_avx2: 286.0
planar_gbrap12be_to_y_512_c: 1701.5
planar_gbrap12be_to_y_512_sse4: 917.0
planar_gbrap12be_to_y_512_avx2: 338.5
planar_gbrap12le_to_y_512_c: 1201.0
planar_gbrap12le_to_y_512_sse4: 444.5
planar_gbrap12le_to_y_512_avx2: 288.0
planar_gbrp14be_to_y_512_c: 1370.5
planar_gbrp14be_to_y_512_sse4: 545.0
planar_gbrp14be_to_y_512_avx2: 338.5
planar_gbrp14le_to_y_512_c: 1199.0
planar_gbrp14le_to_y_512_sse4: 444.0
planar_gbrp14le_to_y_512_avx2: 279.5
planar_gbrp16be_to_y_512_c: 1364.0
planar_gbrp16be_to_y_512_sse4: 544.5
planar_gbrp16be_to_y_512_avx2: 339.5
planar_gbrp16le_to_y_512_c: 1201.0
planar_gbrp16le_to_y_512_sse4: 445.5
planar_gbrp16le_to_y_512_avx2: 280.5
planar_gbrap16be_to_y_512_c: 1377.0
planar_gbrap16be_to_y_512_sse4: 545.0
planar_gbrap16be_to_y_512_avx2: 338.5
planar_gbrap16le_to_y_512_c: 1201.0
planar_gbrap16le_to_y_512_sse4: 442.0
planar_gbrap16le_to_y_512_avx2: 279.0
planar_gbrpf32be_to_y_512_c: 4113.0
planar_gbrpf32be_to_y_512_sse2: 2438.0
planar_gbrpf32be_to_y_512_sse4: 1068.0
planar_gbrpf32be_to_y_512_avx2: 904.5
planar_gbrpf32le_to_y_512_c: 3818.5
planar_gbrpf32le_to_y_512_sse2: 2024.5
planar_gbrpf32le_to_y_512_sse4: 1241.5
planar_gbrpf32le_to_y_512_avx2: 657.0
planar_gbrapf32be_to_y_512_c: 3707.0
planar_gbrapf32be_to_y_512_sse2: 2444.0
planar_gbrapf32be_to_y_512_sse4: 1077.0
planar_gbrapf32be_to_y_512_avx2: 909.0
planar_gbrapf32le_to_y_512_c: 3822.0
planar_gbrapf32le_to_y_512_sse2: 2024.5
planar_gbrapf32le_to_y_512_sse4: 1176.0
planar_gbrapf32le_to_y_512_avx2: 658.5

planar_gbrp_to_uv_512_c: 2325.8
planar_gbrp_to_uv_512_sse2: 1726.8
planar_gbrp_to_uv_512_sse4: 771.8
planar_gbrp_to_uv_512_avx2: 506.8
planar_gbrap_to_uv_512_c: 2281.8
planar_gbrap_to_uv_512_sse2: 1726.3
planar_gbrap_to_uv_512_sse4: 768.3
planar_gbrap_to_uv_512_avx2: 496.3
planar_gbrp9be_to_uv_512_c: 2336.8
planar_gbrp9be_to_uv_512_sse2: 1924.8
planar_gbrp9be_to_uv_512_sse4: 852.3
planar_gbrp9be_to_uv_512_avx2: 552.8
planar_gbrp9le_to_uv_512_c: 2270.3
planar_gbrp9le_to_uv_512_sse2: 1512.3
planar_gbrp9le_to_uv_512_sse4: 764.3
planar_gbrp9le_to_uv_512_avx2: 491.3
planar_gbrp10be_to_uv_512_c: 2281.8
planar_gbrp10be_to_uv_512_sse2: 1917.8
planar_gbrp10be_to_uv_512_sse4: 855.3
planar_gbrp10be_to_uv_512_avx2: 541.3
planar_gbrp10le_to_uv_512_c: 2269.8
planar_gbrp10le_to_uv_512_sse2: 1515.3
planar_gbrp10le_to_uv_512_sse4: 759.8
planar_gbrp10le_to_uv_512_avx2: 487.8
planar_gbrap10be_to_uv_512_c: 2382.3
planar_gbrap10be_to_uv_512_sse2: 1924.8
planar_gbrap10be_to_uv_512_sse4: 855.3
planar_gbrap10be_to_uv_512_avx2: 540.8
planar_gbrap10le_to_uv_512_c: 2382.3
planar_gbrap10le_to_uv_512_sse2: 1512.3
planar_gbrap10le_to_uv_512_sse4: 759.3
planar_gbrap10le_to_uv_512_avx2: 484.8
planar_gbrp12be_to_uv_512_c: 2283.8
planar_gbrp12be_to_uv_512_sse2: 1936.8
planar_gbrp12be_to_uv_512_sse4: 858.3
planar_gbrp12be_to_uv_512_avx2: 541.3
planar_gbrp12le_to_uv_512_c: 2278.8
planar_gbrp12le_to_uv_512_sse2: 1507.3
planar_gbrp12le_to_uv_512_sse4: 760.3
planar_gbrp12le_to_uv_512_avx2: 485.8
planar_gbrap12be_to_uv_512_c: 2385.3
planar_gbrap12be_to_uv_512_sse2: 1927.8
planar_gbrap12be_to_uv_512_sse4: 855.3
planar_gbrap12be_to_uv_512_avx2: 539.8
planar_gbrap12le_to_uv_512_c: 2377.3
planar_gbrap12le_to_uv_512_sse2: 1516.3
planar_gbrap12le_to_uv_512_sse4: 759.3
planar_gbrap12le_to_uv_512_avx2: 484.8
planar_gbrp14be_to_uv_512_c: 2283.8
planar_gbrp14be_to_uv_512_sse2: 1935.3
planar_gbrp14be_to_uv_512_sse4: 852.3
planar_gbrp14be_to_uv_512_avx2: 540.3
planar_gbrp14le_to_uv_512_c: 2276.8
planar_gbrp14le_to_uv_512_sse2: 1514.8
planar_gbrp14le_to_uv_512_sse4: 762.3
planar_gbrp14le_to_uv_512_avx2: 484.8
planar_gbrp16be_to_uv_512_c: 2383.3
planar_gbrp16be_to_uv_512_sse2: 1881.8
planar_gbrp16be_to_uv_512_sse4: 852.3
planar_gbrp16be_to_uv_512_avx2: 541.8
planar_gbrp16le_to_uv_512_c: 2378.3
planar_gbrp16le_to_uv_512_sse2: 1476.8
planar_gbrp16le_to_uv_512_sse4: 765.3
planar_gbrp16le_to_uv_512_avx2: 485.8
planar_gbrap16be_to_uv_512_c: 2382.3
planar_gbrap16be_to_uv_512_sse2: 1886.3
planar_gbrap16be_to_uv_512_sse4: 853.8
planar_gbrap16be_to_uv_512_avx2: 550.8
planar_gbrap16le_to_uv_512_c: 2381.8
planar_gbrap16le_to_uv_512_sse2: 1488.3
planar_gbrap16le_to_uv_512_sse4: 765.3
planar_gbrap16le_to_uv_512_avx2: 491.8
planar_gbrpf32be_to_uv_512_c: 4863.0
planar_gbrpf32be_to_uv_512_sse2: 3347.5
planar_gbrpf32be_to_uv_512_sse4: 1800.0
planar_gbrpf32be_to_uv_512_avx2: 1199.0
planar_gbrpf32le_to_uv_512_c: 4725.0
planar_gbrpf32le_to_uv_512_sse2: 2753.0
planar_gbrpf32le_to_uv_512_sse4: 1474.5
planar_gbrpf32le_to_uv_512_avx2: 927.5
planar_gbrapf32be_to_uv_512_c: 4859.0
planar_gbrapf32be_to_uv_512_sse2: 3269.0
planar_gbrapf32be_to_uv_512_sse4: 1802.0
planar_gbrapf32be_to_uv_512_avx2: 1201.5
planar_gbrapf32le_to_uv_512_c: 6338.0
planar_gbrapf32le_to_uv_512_sse2: 2756.5
planar_gbrapf32le_to_uv_512_sse4: 1476.0
planar_gbrapf32le_to_uv_512_avx2: 908.5

planar_gbrap_to_a_512_c: 383.3
planar_gbrap_to_a_512_sse2: 66.8
planar_gbrap_to_a_512_avx2: 43.8
planar_gbrap10be_to_a_512_c: 601.8
planar_gbrap10be_to_a_512_sse2: 86.3
planar_gbrap10be_to_a_512_avx2: 34.8
planar_gbrap10le_to_a_512_c: 602.3
planar_gbrap10le_to_a_512_sse2: 48.8
planar_gbrap10le_to_a_512_avx2: 31.3
planar_gbrap12be_to_a_512_c: 601.8
planar_gbrap12be_to_a_512_sse2: 111.8
planar_gbrap12be_to_a_512_avx2: 41.3
planar_gbrap12le_to_a_512_c: 385.8
planar_gbrap12le_to_a_512_sse2: 75.3
planar_gbrap12le_to_a_512_avx2: 39.8
planar_gbrap16be_to_a_512_c: 386.8
planar_gbrap16be_to_a_512_sse2: 79.8
planar_gbrap16be_to_a_512_avx2: 31.3
planar_gbrap16le_to_a_512_c: 600.3
planar_gbrap16le_to_a_512_sse2: 40.3
planar_gbrap16le_to_a_512_avx2: 30.3
planar_gbrapf32be_to_a_512_c: 1148.8
planar_gbrapf32be_to_a_512_sse2: 611.3
planar_gbrapf32be_to_a_512_sse4: 234.8
planar_gbrapf32be_to_a_512_avx2: 183.3
planar_gbrapf32le_to_a_512_c: 851.3
planar_gbrapf32le_to_a_512_sse2: 263.3
planar_gbrapf32le_to_a_512_sse4: 199.3
planar_gbrapf32le_to_a_512_avx2: 156.8

Reviewed-by: Paul B Mahol <onemda@gmail.com>
Signed-off-by: James Almer <jamrial@gmail.com>

1 files changed, 216 insertions, 0 deletions

diff --git a/tests/checkasm/sw_gbrp.c b/tests/checkasm/sw_gbrp.c
index 61b2b9d6a1..848e5366ad 100644
--- a/tests/checkasm/sw_gbrp.c
+++ b/tests/checkasm/sw_gbrp.c
@@ -190,9 +190,225 @@ static void check_output_yuv2gbrp(void)
     sws_freeContext(ctx);
 }
 
+#undef LARGEST_INPUT_SIZE
+#undef INPUT_SIZES
+
+static void check_input_planar_rgb_to_y(void)
+{
+    struct SwsContext *ctx;
+    const AVPixFmtDescriptor *desc;
+    int fmi, isi;
+    int dstW, byte_size;
+#define LARGEST_INPUT_SIZE 512
+#define INPUT_SIZES 6
+    static const int input_sizes[] = {8, 24, 128, 144, 256, 512};
+    uint8_t *src[4];
+    int32_t rgb2yuv[9] = {0};
+
+    declare_func(void, uint8_t *dst, uint8_t *src[4], int w, int32_t *rgb2yuv);
+
+    LOCAL_ALIGNED_8(int32_t, src_r, [LARGEST_INPUT_SIZE]);
+    LOCAL_ALIGNED_8(int32_t, src_g, [LARGEST_INPUT_SIZE]);
+    LOCAL_ALIGNED_8(int32_t, src_b, [LARGEST_INPUT_SIZE]);
+    LOCAL_ALIGNED_8(int32_t, src_a, [LARGEST_INPUT_SIZE]);
+
+    LOCAL_ALIGNED_8(uint8_t, dst0_y, [LARGEST_INPUT_SIZE * sizeof(int32_t)]);
+    LOCAL_ALIGNED_8(uint8_t, dst1_y, [LARGEST_INPUT_SIZE * sizeof(int32_t)]);
+
+    randomize_buffers((uint8_t*)src_r, LARGEST_INPUT_SIZE * sizeof(int32_t));
+    randomize_buffers((uint8_t*)src_g, LARGEST_INPUT_SIZE * sizeof(int32_t));
+    randomize_buffers((uint8_t*)src_b, LARGEST_INPUT_SIZE * sizeof(int32_t));
+    randomize_buffers((uint8_t*)src_a, LARGEST_INPUT_SIZE * sizeof(int32_t));
+    randomize_buffers((uint8_t*)rgb2yuv, 9 * sizeof(int32_t));
+
+    src[0] = (uint8_t*)src_g;
+    src[1] = (uint8_t*)src_b;
+    src[2] = (uint8_t*)src_r;
+    src[3] = (uint8_t*)src_a;
+
+    ctx = sws_alloc_context();
+    if (sws_init_context(ctx, NULL, NULL) < 0)
+        fail();
+
+    for (fmi = 0; fmi < FF_ARRAY_ELEMS(planar_fmts); fmi++) {
+        for (isi = 0; isi < INPUT_SIZES; isi++ ) {
+            desc = av_pix_fmt_desc_get(planar_fmts[fmi]);
+            ctx->srcFormat = planar_fmts[fmi];
+            ctx->dstFormat = AV_PIX_FMT_YUVA444P16;
+            byte_size = 2;
+            dstW = input_sizes[isi];
+
+            ff_sws_init_scale(ctx);
+            if(check_func(ctx->readLumPlanar, "planar_%s_to_y_%d",  desc->name, dstW)) {
+                memset(dst0_y, 0xFF, LARGEST_INPUT_SIZE * sizeof(int32_t));
+                memset(dst1_y, 0xFF, LARGEST_INPUT_SIZE * sizeof(int32_t));
+
+                call_ref(dst0_y, src, dstW, rgb2yuv);
+                call_new(dst1_y, src, dstW, rgb2yuv);
+
+                if (memcmp(dst0_y, dst1_y, dstW * byte_size))
+                    fail();
+
+                bench_new(dst1_y, src, dstW, rgb2yuv);
+
+            }
+        }
+    }
+    sws_freeContext(ctx);
+}
+
+#undef LARGEST_INPUT_SIZE
+#undef INPUT_SIZES
+
+static void check_input_planar_rgb_to_uv(void)
+{
+    struct SwsContext *ctx;
+    const AVPixFmtDescriptor *desc;
+    int fmi, isi;
+    int dstW, byte_size;
+#define LARGEST_INPUT_SIZE 512
+#define INPUT_SIZES 6
+    static const int input_sizes[] = {8, 24, 128, 144, 256, 512};
+    uint8_t *src[4];
+    int32_t rgb2yuv[9] = {0};
+
+    declare_func(void, uint8_t *dstU, uint8_t *dstV,
+                       uint8_t *src[4], int w, int32_t *rgb2yuv);
+
+    LOCAL_ALIGNED_8(int32_t, src_r, [LARGEST_INPUT_SIZE]);
+    LOCAL_ALIGNED_8(int32_t, src_g, [LARGEST_INPUT_SIZE]);
+    LOCAL_ALIGNED_8(int32_t, src_b, [LARGEST_INPUT_SIZE]);
+    LOCAL_ALIGNED_8(int32_t, src_a, [LARGEST_INPUT_SIZE]);
+
+    LOCAL_ALIGNED_8(uint8_t, dst0_u, [LARGEST_INPUT_SIZE * sizeof(int32_t)]);
+    LOCAL_ALIGNED_8(uint8_t, dst0_v, [LARGEST_INPUT_SIZE * sizeof(int32_t)]);
+
+    LOCAL_ALIGNED_8(uint8_t, dst1_u, [LARGEST_INPUT_SIZE * sizeof(int32_t)]);
+    LOCAL_ALIGNED_8(uint8_t, dst1_v, [LARGEST_INPUT_SIZE * sizeof(int32_t)]);
+
+    randomize_buffers((uint8_t*)src_r, LARGEST_INPUT_SIZE * sizeof(int32_t));
+    randomize_buffers((uint8_t*)src_g, LARGEST_INPUT_SIZE * sizeof(int32_t));
+    randomize_buffers((uint8_t*)src_b, LARGEST_INPUT_SIZE * sizeof(int32_t));
+    randomize_buffers((uint8_t*)src_a, LARGEST_INPUT_SIZE * sizeof(int32_t));
+    randomize_buffers((uint8_t*)rgb2yuv, 9 * sizeof(int32_t));
+
+    src[0] = (uint8_t*)src_g;
+    src[1] = (uint8_t*)src_b;
+    src[2] = (uint8_t*)src_r;
+    src[3] = (uint8_t*)src_a;
+
+    ctx = sws_alloc_context();
+    if (sws_init_context(ctx, NULL, NULL) < 0)
+        fail();
+
+    for (fmi = 0; fmi < FF_ARRAY_ELEMS(planar_fmts); fmi++) {
+        for (isi = 0; isi < INPUT_SIZES; isi++ ) {
+            desc = av_pix_fmt_desc_get(planar_fmts[fmi]);
+            ctx->srcFormat = planar_fmts[fmi];
+            ctx->dstFormat = AV_PIX_FMT_YUVA444P16;
+            byte_size = 2;
+            dstW = input_sizes[isi];
+
+            ff_sws_init_scale(ctx);
+            if(check_func(ctx->readChrPlanar, "planar_%s_to_uv_%d",  desc->name, dstW)) {
+                memset(dst0_u, 0xFF, LARGEST_INPUT_SIZE * sizeof(int32_t));
+                memset(dst0_v, 0xFF, LARGEST_INPUT_SIZE * sizeof(int32_t));
+                memset(dst1_u, 0xFF, LARGEST_INPUT_SIZE * sizeof(int32_t));
+                memset(dst1_v, 0xFF, LARGEST_INPUT_SIZE * sizeof(int32_t));
+
+                call_ref(dst0_u, dst0_v, src, dstW, rgb2yuv);
+                call_new(dst1_u, dst1_v, src, dstW, rgb2yuv);
+
+                if (memcmp(dst0_u, dst1_u, dstW * byte_size) ||
+                    memcmp(dst0_v, dst1_v, dstW * byte_size))
+                    fail();
+
+                bench_new(dst1_u, dst1_v, src, dstW, rgb2yuv);
+            }
+        }
+    }
+    sws_freeContext(ctx);
+}
+
+#undef LARGEST_INPUT_SIZE
+#undef INPUT_SIZES
+
+static void check_input_planar_rgb_to_a(void)
+{
+    struct SwsContext *ctx;
+    const AVPixFmtDescriptor *desc;
+    int fmi, isi;
+    int dstW, byte_size;
+#define LARGEST_INPUT_SIZE 512
+#define INPUT_SIZES 6
+    static const int input_sizes[] = {8, 24, 128, 144, 256, 512};
+    uint8_t *src[4];
+    int32_t rgb2yuv[9] = {0};
+
+    declare_func(void, uint8_t *dst, uint8_t *src[4], int w, int32_t *rgb2yuv);
+
+    LOCAL_ALIGNED_8(int32_t, src_r, [LARGEST_INPUT_SIZE]);
+    LOCAL_ALIGNED_8(int32_t, src_g, [LARGEST_INPUT_SIZE]);
+    LOCAL_ALIGNED_8(int32_t, src_b, [LARGEST_INPUT_SIZE]);
+    LOCAL_ALIGNED_8(int32_t, src_a, [LARGEST_INPUT_SIZE]);
+
+    LOCAL_ALIGNED_8(uint8_t, dst0_a, [LARGEST_INPUT_SIZE * sizeof(int32_t)]);
+    LOCAL_ALIGNED_8(uint8_t, dst1_a, [LARGEST_INPUT_SIZE * sizeof(int32_t)]);
+
+    randomize_buffers((uint8_t*)src_r, LARGEST_INPUT_SIZE * sizeof(int32_t));
+    randomize_buffers((uint8_t*)src_g, LARGEST_INPUT_SIZE * sizeof(int32_t));
+    randomize_buffers((uint8_t*)src_b, LARGEST_INPUT_SIZE * sizeof(int32_t));
+    randomize_buffers((uint8_t*)src_a, LARGEST_INPUT_SIZE * sizeof(int32_t));
+    randomize_buffers((uint8_t*)rgb2yuv, 9 * sizeof(int32_t));
+
+    src[0] = (uint8_t*)src_g;
+    src[1] = (uint8_t*)src_b;
+    src[2] = (uint8_t*)src_r;
+    src[3] = (uint8_t*)src_a;
+
+    ctx = sws_alloc_context();
+    if (sws_init_context(ctx, NULL, NULL) < 0)
+        fail();
+
+    for (fmi = 0; fmi < FF_ARRAY_ELEMS(planar_fmts); fmi++) {
+        for (isi = 0; isi < INPUT_SIZES; isi++ ) {
+            desc = av_pix_fmt_desc_get(planar_fmts[fmi]);
+            if (!(desc->flags & AV_PIX_FMT_FLAG_ALPHA))
+                continue;
+
+            ctx->srcFormat = planar_fmts[fmi];
+            ctx->dstFormat = AV_PIX_FMT_YUVA444P16;
+            byte_size = 2;
+            dstW = input_sizes[isi];
+
+            ff_sws_init_scale(ctx);
+            if(check_func(ctx->readAlpPlanar, "planar_%s_to_a_%d",  desc->name, dstW)) {
+                memset(dst0_a, 0x00, LARGEST_INPUT_SIZE * sizeof(int32_t));
+                memset(dst1_a, 0x00, LARGEST_INPUT_SIZE * sizeof(int32_t));
+
+                call_ref(dst0_a, src, dstW, rgb2yuv);
+                call_new(dst1_a, src, dstW, rgb2yuv);
+
+                if (memcmp(dst0_a, dst1_a, dstW * byte_size))
+                    fail();
+                bench_new(dst1_a, src, dstW, rgb2yuv);
+            }
+        }
+    }
+    sws_freeContext(ctx);
+}
 
 void checkasm_check_sw_gbrp(void)
 {
     check_output_yuv2gbrp();
     report("output_yuv2gbrp");
+
+    check_input_planar_rgb_to_y();
+    report("input_planar_rgb_y");
+
+    check_input_planar_rgb_to_uv();
+    report("input_planar_rgb_uv");
+
+    check_input_planar_rgb_to_a();
+    report("input_planar_rgb_a");
 }