diff --git a/src/dsp/lossless_mips_dsp_r2.c b/src/dsp/lossless_mips_dsp_r2.c index 02d49595..849142b0 100644 --- a/src/dsp/lossless_mips_dsp_r2.c +++ b/src/dsp/lossless_mips_dsp_r2.c @@ -145,6 +145,61 @@ static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1, return temp1; } +static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, + int num_pixels) { + uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; + uint32_t* const p_loop1_end = argb_data + (num_pixels & ~3); + uint32_t* const p_loop2_end = p_loop1_end + (num_pixels & 3); + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "beq %[argb_data], %[p_loop1_end], 3f \n\t" + " nop \n\t" + "0: \n\t" + "lw %[temp0], 0(%[argb_data]) \n\t" + "lw %[temp1], 4(%[argb_data]) \n\t" + "lw %[temp2], 8(%[argb_data]) \n\t" + "lw %[temp3], 12(%[argb_data]) \n\t" + "ext %[temp4], %[temp0], 8, 8 \n\t" + "ext %[temp5], %[temp1], 8, 8 \n\t" + "ext %[temp6], %[temp2], 8, 8 \n\t" + "ext %[temp7], %[temp3], 8, 8 \n\t" + "addiu %[argb_data], %[argb_data], 16 \n\t" + "replv.ph %[temp4], %[temp4] \n\t" + "replv.ph %[temp5], %[temp5] \n\t" + "replv.ph %[temp6], %[temp6] \n\t" + "replv.ph %[temp7], %[temp7] \n\t" + "subu.qb %[temp0], %[temp0], %[temp4] \n\t" + "subu.qb %[temp1], %[temp1], %[temp5] \n\t" + "subu.qb %[temp2], %[temp2], %[temp6] \n\t" + "subu.qb %[temp3], %[temp3], %[temp7] \n\t" + "sw %[temp0], -16(%[argb_data]) \n\t" + "sw %[temp1], -12(%[argb_data]) \n\t" + "sw %[temp2], -8(%[argb_data]) \n\t" + "bne %[argb_data], %[p_loop1_end], 0b \n\t" + " sw %[temp3], -4(%[argb_data]) \n\t" + "3: \n\t" + "beq %[argb_data], %[p_loop2_end], 2f \n\t" + " nop \n\t" + "1: \n\t" + "lw %[temp0], 0(%[argb_data]) \n\t" + "addiu %[argb_data], %[argb_data], 4 \n\t" + "ext %[temp4], %[temp0], 8, 8 \n\t" + "replv.ph %[temp4], %[temp4] \n\t" + "subu.qb %[temp0], %[temp0], %[temp4] \n\t" + "bne %[argb_data], %[p_loop2_end], 1b \n\t" + " sw %[temp0], -4(%[argb_data]) \n\t" + "2: \n\t" + ".set pop \n\t" + : [argb_data]"+&r"(argb_data), [temp0]"=&r"(temp0), + [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [temp6]"=&r"(temp6), + [temp7]"=&r"(temp7) + : [p_loop1_end]"r"(p_loop1_end), [p_loop2_end]"r"(p_loop2_end) + : "memory" + ); +} + static uint32_t Predictor12(uint32_t left, const uint32_t* const top) { return ClampedAddSubtractFull(left, top[0], top[-1]); } @@ -165,6 +220,7 @@ void VP8LDspInitMIPSdspR2(void) { VP8LMapColor8b = MapAlpha; VP8LPredictors[12] = Predictor12; VP8LPredictors[13] = Predictor13; + VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed; #endif // WEBP_USE_MIPS_DSP_R2 }