diff --git a/src/dsp/lossless.h b/src/dsp/lossless.h index 280e560a..4d017160 100644 --- a/src/dsp/lossless.h +++ b/src/dsp/lossless.h @@ -35,6 +35,7 @@ extern "C" { typedef uint32_t (*VP8LPredictorFunc)(uint32_t left, const uint32_t* const top); extern VP8LPredictorFunc VP8LPredictors[16]; extern VP8LPredictorFunc VP8LPredictors_C[16]; +// These Add/Sub function expects upper[-1] and out[-1] to be readable. typedef void (*VP8LPredictorAddSubFunc)(const uint32_t* in, const uint32_t* upper, int num_pixels, uint32_t* out); @@ -150,6 +151,7 @@ void VP8LCollectColorBlueTransforms_C(const uint32_t* argb, int stride, int histo[]); extern VP8LPredictorAddSubFunc VP8LPredictorsSub[16]; +extern VP8LPredictorAddSubFunc VP8LPredictorsSub_C[16]; // ----------------------------------------------------------------------------- // Huffman-cost related functions. diff --git a/src/dsp/lossless_enc.c b/src/dsp/lossless_enc.c index 32ad5413..356f3b82 100644 --- a/src/dsp/lossless_enc.c +++ b/src/dsp/lossless_enc.c @@ -663,6 +663,141 @@ static void HistogramAdd(const VP8LHistogram* const a, } } +//------------------------------------------------------------------------------ +// Image transforms. + +static WEBP_INLINE uint32_t Average2(uint32_t a0, uint32_t a1) { + return (((a0 ^ a1) & 0xfefefefeu) >> 1) + (a0 & a1); +} + +static WEBP_INLINE uint32_t Average3(uint32_t a0, uint32_t a1, uint32_t a2) { + return Average2(Average2(a0, a2), a1); +} + +static WEBP_INLINE uint32_t Average4(uint32_t a0, uint32_t a1, + uint32_t a2, uint32_t a3) { + return Average2(Average2(a0, a1), Average2(a2, a3)); +} + +static WEBP_INLINE uint32_t Clip255(uint32_t a) { + if (a < 256) { + return a; + } + // return 0, when a is a negative integer. + // return 255, when a is positive. + return ~a >> 24; +} + +static WEBP_INLINE int AddSubtractComponentFull(int a, int b, int c) { + return Clip255(a + b - c); +} + +static WEBP_INLINE uint32_t ClampedAddSubtractFull(uint32_t c0, uint32_t c1, + uint32_t c2) { + const int a = AddSubtractComponentFull(c0 >> 24, c1 >> 24, c2 >> 24); + const int r = AddSubtractComponentFull((c0 >> 16) & 0xff, + (c1 >> 16) & 0xff, + (c2 >> 16) & 0xff); + const int g = AddSubtractComponentFull((c0 >> 8) & 0xff, + (c1 >> 8) & 0xff, + (c2 >> 8) & 0xff); + const int b = AddSubtractComponentFull(c0 & 0xff, c1 & 0xff, c2 & 0xff); + return ((uint32_t)a << 24) | (r << 16) | (g << 8) | b; +} + +static WEBP_INLINE int AddSubtractComponentHalf(int a, int b) { + return Clip255(a + (a - b) / 2); +} + +static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1, + uint32_t c2) { + const uint32_t ave = Average2(c0, c1); + const int a = AddSubtractComponentHalf(ave >> 24, c2 >> 24); + const int r = AddSubtractComponentHalf((ave >> 16) & 0xff, (c2 >> 16) & 0xff); + const int g = AddSubtractComponentHalf((ave >> 8) & 0xff, (c2 >> 8) & 0xff); + const int b = AddSubtractComponentHalf((ave >> 0) & 0xff, (c2 >> 0) & 0xff); + return ((uint32_t)a << 24) | (r << 16) | (g << 8) | b; +} + +// gcc-4.9 on ARM generates incorrect code in Select() when Sub3() is inlined. +#if defined(__arm__) && \ + (LOCAL_GCC_VERSION == 0x409 || LOCAL_GCC_VERSION == 0x408) +# define LOCAL_INLINE __attribute__ ((noinline)) +#else +# define LOCAL_INLINE WEBP_INLINE +#endif + +static LOCAL_INLINE int Sub3(int a, int b, int c) { + const int pb = b - c; + const int pa = a - c; + return abs(pb) - abs(pa); +} + +#undef LOCAL_INLINE + +static WEBP_INLINE uint32_t Select(uint32_t a, uint32_t b, uint32_t c) { + const int pa_minus_pb = + Sub3((a >> 24) , (b >> 24) , (c >> 24) ) + + Sub3((a >> 16) & 0xff, (b >> 16) & 0xff, (c >> 16) & 0xff) + + Sub3((a >> 8) & 0xff, (b >> 8) & 0xff, (c >> 8) & 0xff) + + Sub3((a ) & 0xff, (b ) & 0xff, (c ) & 0xff); + return (pa_minus_pb <= 0) ? a : b; +} + +//------------------------------------------------------------------------------ +// Predictors + +static uint32_t Predictor2(uint32_t left, const uint32_t* const top) { + (void)left; + return top[0]; +} +static uint32_t Predictor3(uint32_t left, const uint32_t* const top) { + (void)left; + return top[1]; +} +static uint32_t Predictor4(uint32_t left, const uint32_t* const top) { + (void)left; + return top[-1]; +} +static uint32_t Predictor5(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Average3(left, top[0], top[1]); + return pred; +} +static uint32_t Predictor6(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Average2(left, top[-1]); + return pred; +} +static uint32_t Predictor7(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Average2(left, top[0]); + return pred; +} +static uint32_t Predictor8(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Average2(top[-1], top[0]); + (void)left; + return pred; +} +static uint32_t Predictor9(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Average2(top[0], top[1]); + (void)left; + return pred; +} +static uint32_t Predictor10(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Average4(left, top[-1], top[0], top[1]); + return pred; +} +static uint32_t Predictor11(uint32_t left, const uint32_t* const top) { + const uint32_t pred = Select(top[0], left, top[-1]); + return pred; +} +static uint32_t Predictor12(uint32_t left, const uint32_t* const top) { + const uint32_t pred = ClampedAddSubtractFull(left, top[0], top[-1]); + return pred; +} +static uint32_t Predictor13(uint32_t left, const uint32_t* const top) { + const uint32_t pred = ClampedAddSubtractHalf(left, top[0], top[-1]); + return pred; +} + //------------------------------------------------------------------------------ static void PredictorSub0_C(const uint32_t* in, const uint32_t* upper, @@ -679,18 +814,20 @@ static void PredictorSub1_C(const uint32_t* in, const uint32_t* upper, (void)upper; } -GENERATE_PREDICTOR_SUB(VP8LPredictors[2], PredictorSub2_C) -GENERATE_PREDICTOR_SUB(VP8LPredictors[3], PredictorSub3_C) -GENERATE_PREDICTOR_SUB(VP8LPredictors[4], PredictorSub4_C) -GENERATE_PREDICTOR_SUB(VP8LPredictors[5], PredictorSub5_C) -GENERATE_PREDICTOR_SUB(VP8LPredictors[6], PredictorSub6_C) -GENERATE_PREDICTOR_SUB(VP8LPredictors[7], PredictorSub7_C) -GENERATE_PREDICTOR_SUB(VP8LPredictors[8], PredictorSub8_C) -GENERATE_PREDICTOR_SUB(VP8LPredictors[9], PredictorSub9_C) -GENERATE_PREDICTOR_SUB(VP8LPredictors[10], PredictorSub10_C) -GENERATE_PREDICTOR_SUB(VP8LPredictors[11], PredictorSub11_C) -GENERATE_PREDICTOR_SUB(VP8LPredictors[12], PredictorSub12_C) -GENERATE_PREDICTOR_SUB(VP8LPredictors[13], PredictorSub13_C) +GENERATE_PREDICTOR_SUB(Predictor2, PredictorSub2_C) +GENERATE_PREDICTOR_SUB(Predictor3, PredictorSub3_C) +GENERATE_PREDICTOR_SUB(Predictor4, PredictorSub4_C) +GENERATE_PREDICTOR_SUB(Predictor5, PredictorSub5_C) +GENERATE_PREDICTOR_SUB(Predictor6, PredictorSub6_C) +GENERATE_PREDICTOR_SUB(Predictor7, PredictorSub7_C) +GENERATE_PREDICTOR_SUB(Predictor8, PredictorSub8_C) +GENERATE_PREDICTOR_SUB(Predictor9, PredictorSub9_C) +GENERATE_PREDICTOR_SUB(Predictor10, PredictorSub10_C) +GENERATE_PREDICTOR_SUB(Predictor11, PredictorSub11_C) +GENERATE_PREDICTOR_SUB(Predictor12, PredictorSub12_C) +GENERATE_PREDICTOR_SUB(Predictor13, PredictorSub13_C) + +//------------------------------------------------------------------------------ VP8LProcessEncBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed; @@ -714,6 +851,7 @@ VP8LHistogramAddFunc VP8LHistogramAdd; VP8LVectorMismatchFunc VP8LVectorMismatch; VP8LPredictorAddSubFunc VP8LPredictorsSub[16]; +VP8LPredictorAddSubFunc VP8LPredictorsSub_C[16]; extern void VP8LEncDspInitSSE2(void); extern void VP8LEncDspInitSSE41(void); @@ -768,6 +906,23 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInit(void) { VP8LPredictorsSub[14] = PredictorSub0_C; // <- padding security sentinels VP8LPredictorsSub[15] = PredictorSub0_C; + VP8LPredictorsSub_C[0] = PredictorSub0_C; + VP8LPredictorsSub_C[1] = PredictorSub1_C; + VP8LPredictorsSub_C[2] = PredictorSub2_C; + VP8LPredictorsSub_C[3] = PredictorSub3_C; + VP8LPredictorsSub_C[4] = PredictorSub4_C; + VP8LPredictorsSub_C[5] = PredictorSub5_C; + VP8LPredictorsSub_C[6] = PredictorSub6_C; + VP8LPredictorsSub_C[7] = PredictorSub7_C; + VP8LPredictorsSub_C[8] = PredictorSub8_C; + VP8LPredictorsSub_C[9] = PredictorSub9_C; + VP8LPredictorsSub_C[10] = PredictorSub10_C; + VP8LPredictorsSub_C[11] = PredictorSub11_C; + VP8LPredictorsSub_C[12] = PredictorSub12_C; + VP8LPredictorsSub_C[13] = PredictorSub13_C; + VP8LPredictorsSub_C[14] = PredictorSub0_C; // <- padding security sentinels + VP8LPredictorsSub_C[15] = PredictorSub0_C; + // If defined, use CPUInfo() to overwrite some pointers with faster versions. if (VP8GetCPUInfo != NULL) { #if defined(WEBP_USE_SSE2) diff --git a/src/dsp/lossless_enc_sse2.c b/src/dsp/lossless_enc_sse2.c index 46a55a2f..805f6860 100644 --- a/src/dsp/lossless_enc_sse2.c +++ b/src/dsp/lossless_enc_sse2.c @@ -377,6 +377,203 @@ static int VectorMismatch(const uint32_t* const array1, return match_len; } +//------------------------------------------------------------------------------ +// Batch version of Predictor Transform subtraction + +static WEBP_INLINE void Average2_m128i(const __m128i* const a0, + const __m128i* const a1, + __m128i* const avg) { + // (a + b) >> 1 = ((a + b + 1) >> 1) - ((a ^ b) & 1) + const __m128i ones = _mm_set1_epi8(1); + const __m128i avg1 = _mm_avg_epu8(*a0, *a1); + const __m128i one = _mm_and_si128(_mm_xor_si128(*a0, *a1), ones); + *avg = _mm_sub_epi8(avg1, one); +} + +// Predictor0: ARGB_BLACK. +static void PredictorSub0_SSE2(const uint32_t* in, const uint32_t* upper, + int num_pixels, uint32_t* out) { + int i; + const __m128i black = _mm_set1_epi32(ARGB_BLACK); + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); + const __m128i res = _mm_sub_epi8(src, black); + _mm_storeu_si128((__m128i*)&out[i], res); + } + VP8LPredictorsSub_C[0](in + i, upper + i, num_pixels - i, out + i); +} + +#define GENERATE_PREDICTOR_1(X, IN) \ +static void PredictorSub##X##_SSE2(const uint32_t* in, const uint32_t* upper, \ + int num_pixels, uint32_t* out) { \ + int i; \ + for (i = 0; i + 4 <= num_pixels; i += 4) { \ + const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \ + const __m128i pred = _mm_loadu_si128((const __m128i*)&(IN)); \ + const __m128i res = _mm_sub_epi8(src, pred); \ + _mm_storeu_si128((__m128i*)&out[i], res); \ + } \ + VP8LPredictorsSub_C[(X)](in + i, upper + i, num_pixels - i, out + i); \ +} + +GENERATE_PREDICTOR_1(1, in[i - 1]) // Predictor1: L +GENERATE_PREDICTOR_1(2, upper[i]) // Predictor2: T +GENERATE_PREDICTOR_1(3, upper[i + 1]) // Predictor3: TR +GENERATE_PREDICTOR_1(4, upper[i - 1]) // Predictor4: TL +#undef GENERATE_PREDICTOR_1 + +// Predictor5: avg2(avg2(L, TR), T) +static void PredictorSub5_SSE2(const uint32_t* in, const uint32_t* upper, + int num_pixels, uint32_t* out) { + int i; + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]); + const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); + const __m128i TR = _mm_loadu_si128((const __m128i*)&upper[i + 1]); + const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); + __m128i avg, pred, res; + Average2_m128i(&L, &TR, &avg); + Average2_m128i(&avg, &T, &pred); + res = _mm_sub_epi8(src, pred); + _mm_storeu_si128((__m128i*)&out[i], res); + } + VP8LPredictorsSub_C[5](in + i, upper + i, num_pixels - i, out + i); +} + +#define GENERATE_PREDICTOR_2(X, A, B) \ +static void PredictorSub##X##_SSE2(const uint32_t* in, const uint32_t* upper, \ + int num_pixels, uint32_t* out) { \ + int i; \ + for (i = 0; i + 4 <= num_pixels; i += 4) { \ + const __m128i tA = _mm_loadu_si128((const __m128i*)&(A)); \ + const __m128i tB = _mm_loadu_si128((const __m128i*)&(B)); \ + const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \ + __m128i pred, res; \ + Average2_m128i(&tA, &tB, &pred); \ + res = _mm_sub_epi8(src, pred); \ + _mm_storeu_si128((__m128i*)&out[i], res); \ + } \ + VP8LPredictorsSub_C[(X)](in + i, upper + i, num_pixels - i, out + i); \ +} + +GENERATE_PREDICTOR_2(6, in[i - 1], upper[i - 1]) // Predictor6: avg(L, TL) +GENERATE_PREDICTOR_2(7, in[i - 1], upper[i]) // Predictor7: avg(L, T) +GENERATE_PREDICTOR_2(8, upper[i - 1], upper[i]) // Predictor8: avg(TL, T) +GENERATE_PREDICTOR_2(9, upper[i], upper[i + 1]) // Predictor9: average(T, TR) +#undef GENERATE_PREDICTOR_2 + +// Predictor10: avg(avg(L,TL), avg(T, TR)). +static void PredictorSub10_SSE2(const uint32_t* in, const uint32_t* upper, + int num_pixels, uint32_t* out) { + int i; + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]); + const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); + const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]); + const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); + const __m128i TR = _mm_loadu_si128((const __m128i*)&upper[i + 1]); + __m128i avgTTR, avgLTL, avg, res; + Average2_m128i(&T, &TR, &avgTTR); + Average2_m128i(&L, &TL, &avgLTL); + Average2_m128i(&avgTTR, &avgLTL, &avg); + res = _mm_sub_epi8(src, avg); + _mm_storeu_si128((__m128i*)&out[i], res); + } + VP8LPredictorsSub_C[10](in + i, upper + i, num_pixels - i, out + i); +} + +// Predictor11: select. +static void GetSumAbsDiff32(const __m128i* const A, const __m128i* const B, + __m128i* const out) { + // We can unpack with any value on the upper 32 bits, provided it's the same + // on both operands (to that their sum of abs diff is zero). Here we use *A. + const __m128i A_lo = _mm_unpacklo_epi32(*A, *A); + const __m128i B_lo = _mm_unpacklo_epi32(*B, *A); + const __m128i A_hi = _mm_unpackhi_epi32(*A, *A); + const __m128i B_hi = _mm_unpackhi_epi32(*B, *A); + const __m128i s_lo = _mm_sad_epu8(A_lo, B_lo); + const __m128i s_hi = _mm_sad_epu8(A_hi, B_hi); + *out = _mm_packs_epi32(s_lo, s_hi); +} + +static void PredictorSub11_SSE2(const uint32_t* in, const uint32_t* upper, + int num_pixels, uint32_t* out) { + int i; + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]); + const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); + const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]); + const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); + __m128i pa, pb; + GetSumAbsDiff32(&T, &TL, &pa); // pa = sum |T-TL| + GetSumAbsDiff32(&L, &TL, &pb); // pb = sum |L-TL| + { + const __m128i mask = _mm_cmpgt_epi32(pb, pa); + const __m128i A = _mm_and_si128(mask, L); + const __m128i B = _mm_andnot_si128(mask, T); + const __m128i pred = _mm_or_si128(A, B); // pred = (L > T)? L : T + const __m128i res = _mm_sub_epi8(src, pred); + _mm_storeu_si128((__m128i*)&out[i], res); + } + } + VP8LPredictorsSub_C[11](in + i, upper + i, num_pixels - i, out + i); +} + +// Predictor12: ClampedSubSubtractFull. +static void PredictorSub12_SSE2(const uint32_t* in, const uint32_t* upper, + int num_pixels, uint32_t* out) { + int i; + const __m128i zero = _mm_setzero_si128(); + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); + const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]); + const __m128i L_lo = _mm_unpacklo_epi8(L, zero); + const __m128i L_hi = _mm_unpackhi_epi8(L, zero); + const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); + const __m128i T_lo = _mm_unpacklo_epi8(T, zero); + const __m128i T_hi = _mm_unpackhi_epi8(T, zero); + const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]); + const __m128i TL_lo = _mm_unpacklo_epi8(TL, zero); + const __m128i TL_hi = _mm_unpackhi_epi8(TL, zero); + const __m128i diff_lo = _mm_sub_epi16(T_lo, TL_lo); + const __m128i diff_hi = _mm_sub_epi16(T_hi, TL_hi); + const __m128i pred_lo = _mm_add_epi16(L_lo, diff_lo); + const __m128i pred_hi = _mm_add_epi16(L_hi, diff_hi); + const __m128i pred = _mm_packus_epi16(pred_lo, pred_hi); + const __m128i res = _mm_sub_epi8(src, pred); + _mm_storeu_si128((__m128i*)&out[i], res); + } + VP8LPredictorsSub_C[12](in + i, upper + i, num_pixels - i, out + i); +} + +// Predictors13: ClampedAddSubtractHalf +static void PredictorSub13_SSE2(const uint32_t* in, const uint32_t* upper, + int num_pixels, uint32_t* out) { + int i; + const __m128i zero = _mm_setzero_si128(); + for (i = 0; i + 2 <= num_pixels; i += 2) { + // we can only process two pixels at a time + const __m128i L = _mm_loadl_epi64((const __m128i*)&in[i - 1]); + const __m128i src = _mm_loadl_epi64((const __m128i*)&in[i]); + const __m128i T = _mm_loadl_epi64((const __m128i*)&upper[i]); + const __m128i TL = _mm_loadl_epi64((const __m128i*)&upper[i - 1]); + const __m128i L_lo = _mm_unpacklo_epi8(L, zero); + const __m128i T_lo = _mm_unpacklo_epi8(T, zero); + const __m128i TL_lo = _mm_unpacklo_epi8(TL, zero); + const __m128i sum = _mm_add_epi16(T_lo, L_lo); + const __m128i avg = _mm_srli_epi16(sum, 1); + const __m128i A1 = _mm_sub_epi16(avg, TL_lo); + const __m128i bit_fix = _mm_cmpgt_epi16(TL_lo, avg); + const __m128i A2 = _mm_sub_epi16(A1, bit_fix); + const __m128i A3 = _mm_srai_epi16(A2, 1); + const __m128i A4 = _mm_add_epi16(avg, A3); + const __m128i pred = _mm_packus_epi16(A4, A4); + const __m128i res = _mm_sub_epi8(src, pred); + _mm_storel_epi64((__m128i*)&out[i], res); + } + VP8LPredictorsSub_C[13](in + i, upper + i, num_pixels - i, out + i); +} + //------------------------------------------------------------------------------ // Entry point @@ -390,6 +587,23 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE2(void) { VP8LHistogramAdd = HistogramAdd; VP8LCombinedShannonEntropy = CombinedShannonEntropy; VP8LVectorMismatch = VectorMismatch; + + VP8LPredictorsSub[0] = PredictorSub0_SSE2; + VP8LPredictorsSub[1] = PredictorSub1_SSE2; + VP8LPredictorsSub[2] = PredictorSub2_SSE2; + VP8LPredictorsSub[3] = PredictorSub3_SSE2; + VP8LPredictorsSub[4] = PredictorSub4_SSE2; + VP8LPredictorsSub[5] = PredictorSub5_SSE2; + VP8LPredictorsSub[6] = PredictorSub6_SSE2; + VP8LPredictorsSub[7] = PredictorSub7_SSE2; + VP8LPredictorsSub[8] = PredictorSub8_SSE2; + VP8LPredictorsSub[9] = PredictorSub9_SSE2; + VP8LPredictorsSub[10] = PredictorSub10_SSE2; + VP8LPredictorsSub[11] = PredictorSub11_SSE2; + VP8LPredictorsSub[12] = PredictorSub12_SSE2; + VP8LPredictorsSub[13] = PredictorSub13_SSE2; + VP8LPredictorsSub[14] = PredictorSub0_SSE2; // <- padding security sentinels + VP8LPredictorsSub[15] = PredictorSub0_SSE2; } #else // !WEBP_USE_SSE2