From: Ronald S. Bultje [mailto:rsbul...@gmail.com] Sent: 16 July 2015 20:49 To: FFmpeg development discussions and patches Cc: Rob Isherwood; Shivraj Patil Subject: Re: [FFmpeg-devel] [PATCH 4/4] avcodec/mips: MSA (MIPS-SIMD-Arch) optimizations for VP9 intra functions
Hi, On Thu, Jul 9, 2015 at 9:15 AM, <shivraj.pa...@imgtec.com<mailto:shivraj.pa...@imgtec.com>> wrote: +static void intra_predict_vert_4x4_msa(const uint8_t *src, uint8_t *dst, + int32_t dst_stride) +{ + uint32_t src_data; + + src_data = LW(src); + + SW4(src_data, src_data, src_data, src_data, dst, dst_stride); +} Is this faster than the C function? I know this is a fair bit of work, but ideally you'd profile each individual simd function to see how much faster it is than the C function. These won't be faster, so they just increase the binary size. Same is likely true for e.g. the vert_8x8 one. Shivraj:-Yes, there is no significant gain, so will keep the original c functions for both 4x4, 8x8 horiz and vert cases. +static void intra_predict_horiz_4x4_msa(const uint8_t *src, int32_t src_stride, + uint8_t *dst, int32_t dst_stride) +{ + uint32_t out0, out1, out2, out3; + + out0 = src[0 * src_stride] * 0x01010101; + out1 = src[1 * src_stride] * 0x01010101; + out2 = src[2 * src_stride] * 0x01010101; + out3 = src[3 * src_stride] * 0x01010101; + + SW4(out0, out1, out2, out3, dst, dst_stride); +} Same question here - I suspect this isn't faster than the C version. Same for horiz_8x8. +static void intra_predict_dc_4x4_msa(const uint8_t *src_top, + const uint8_t *src_left, + int32_t src_stride_left, + uint8_t *dst, int32_t dst_stride, + uint8_t is_above, uint8_t is_left) +{ + uint32_t row; + uint32_t out, addition = 0; + v16u8 src_above, store; + v8u16 sum_above; + v4u32 sum; + + if (is_left && is_above) { + src_above = LD_UB(src_top); + + sum_above = __msa_hadd_u_h(src_above, src_above); + sum = __msa_hadd_u_w(sum_above, sum_above); + addition = __msa_copy_u_w((v4i32) sum, 0); + + for (row = 0; row < 4; row++) { + addition += src_left[row * src_stride_left]; + } + + addition = (addition + 4) >> 3; + store = (v16u8) __msa_fill_b(addition); + } else if (is_left) { + for (row = 0; row < 4; row++) { + addition += src_left[row * src_stride_left]; + } + + addition = (addition + 2) >> 2; + store = (v16u8) __msa_fill_b(addition); + } else if (is_above) { + src_above = LD_UB(src_top); + + sum_above = __msa_hadd_u_h(src_above, src_above); + sum = __msa_hadd_u_w(sum_above, sum_above); + sum = (v4u32) __msa_srari_w((v4i32) sum, 2); + store = (v16u8) __msa_splati_b((v16i8) sum, 0); + } else { + store = (v16u8) __msa_ldi_b(128); + } + + out = __msa_copy_u_w((v4i32) store, 0); + + for (row = 4; row--;) { + SW(out, dst); + dst += dst_stride; + } +} + +static void intra_predict_dc_8x8_msa(const uint8_t *src_top, + const uint8_t *src_left, + int32_t src_stride_left, + uint8_t *dst, int32_t dst_stride, + uint8_t is_above, uint8_t is_left) +{ + uint32_t row; + uint32_t out, addition = 0; + v16u8 src_above, store; + v8u16 sum_above; + v4u32 sum_top; + v2u64 sum; + + if (is_left && is_above) { + src_above = LD_UB(src_top); + + sum_above = __msa_hadd_u_h(src_above, src_above); + sum_top = __msa_hadd_u_w(sum_above, sum_above); + sum = __msa_hadd_u_d(sum_top, sum_top); + addition = __msa_copy_u_w((v4i32) sum, 0); + + for (row = 0; row < 8; row++) { + addition += src_left[row * src_stride_left]; + } + + addition = (addition + 8) >> 4; + store = (v16u8) __msa_fill_b(addition); + } else if (is_left) { + for (row = 0; row < 8; row++) { + addition += src_left[row * src_stride_left]; + } + + addition = (addition + 4) >> 3; + store = (v16u8) __msa_fill_b(addition); + } else if (is_above) { + src_above = LD_UB(src_top); + + sum_above = __msa_hadd_u_h(src_above, src_above); + sum_top = __msa_hadd_u_w(sum_above, sum_above); + sum = __msa_hadd_u_d(sum_top, sum_top); + sum = (v2u64) __msa_srari_d((v2i64) sum, 3); + store = (v16u8) __msa_splati_b((v16i8) sum, 0); + } else { + store = (v16u8) __msa_ldi_b(128); + } + + out = __msa_copy_u_w((v4i32) store, 0); + + for (row = 8; row--;) { + SW(out, dst); + SW(out, (dst + 4)); + dst += dst_stride; + } +} + +static void intra_predict_dc_16x16_msa(const uint8_t *src_top, + const uint8_t *src_left, + int32_t src_stride_left, + uint8_t *dst, int32_t dst_stride, + uint8_t is_above, uint8_t is_left) +{ + uint32_t row; + uint32_t addition = 0; + v16u8 src_above, store; + v8u16 sum_above; + v4u32 sum_top; + v2u64 sum; + + if (is_left && is_above) { + src_above = LD_UB(src_top); + + sum_above = __msa_hadd_u_h(src_above, src_above); + sum_top = __msa_hadd_u_w(sum_above, sum_above); + sum = __msa_hadd_u_d(sum_top, sum_top); + sum_top = (v4u32) __msa_pckev_w((v4i32) sum, (v4i32) sum); + sum = __msa_hadd_u_d(sum_top, sum_top); + addition = __msa_copy_u_w((v4i32) sum, 0); + + for (row = 0; row < 16; row++) { + addition += src_left[row * src_stride_left]; + } + + addition = (addition + 16) >> 5; + store = (v16u8) __msa_fill_b(addition); + } else if (is_left) { + for (row = 0; row < 16; row++) { + addition += src_left[row * src_stride_left]; + } + + addition = (addition + 8) >> 4; + store = (v16u8) __msa_fill_b(addition); + } else if (is_above) { + src_above = LD_UB(src_top); + + sum_above = __msa_hadd_u_h(src_above, src_above); + sum_top = __msa_hadd_u_w(sum_above, sum_above); + sum = __msa_hadd_u_d(sum_top, sum_top); + sum_top = (v4u32) __msa_pckev_w((v4i32) sum, (v4i32) sum); + sum = __msa_hadd_u_d(sum_top, sum_top); + sum = (v2u64) __msa_srari_d((v2i64) sum, 4); + store = (v16u8) __msa_splati_b((v16i8) sum, 0); + } else { + store = (v16u8) __msa_ldi_b(128); + } + + for (row = 16; row--;) { + ST_UB(store, dst); + dst += dst_stride; + } +} + +static void intra_predict_dc_32x32_msa(const uint8_t *src_top, + const uint8_t *src_left, + int32_t src_stride_left, + uint8_t *dst, int32_t dst_stride, + uint8_t is_above, uint8_t is_left) +{ + uint32_t row; + uint32_t addition = 0; + v16u8 src_above1, src_above2, store; + v8u16 sum_above1, sum_above2, sum_above; + v4u32 sum_top; + v2u64 sum; + + if (is_left && is_above) { + src_above1 = LD_UB(src_top); + src_above2 = LD_UB(src_top + 16); + + HADD_UB2_UH(src_above1, src_above2, sum_above1, sum_above2); + + sum_above = sum_above1 + sum_above2; + sum_top = __msa_hadd_u_w(sum_above, sum_above); + sum = __msa_hadd_u_d(sum_top, sum_top); + sum_top = (v4u32) __msa_pckev_w((v4i32) sum, (v4i32) sum); + sum = __msa_hadd_u_d(sum_top, sum_top); + addition = __msa_copy_u_w((v4i32) sum, 0); + + for (row = 0; row < 32; row++) { + addition += src_left[row * src_stride_left]; + } + + addition = (addition + 32) >> 6; + store = (v16u8) __msa_fill_b(addition); + } else if (is_left) { + for (row = 0; row < 32; row++) { + addition += src_left[row * src_stride_left]; + } + + addition = (addition + 16) >> 5; + store = (v16u8) __msa_fill_b(addition); + } else if (is_above) { + src_above1 = LD_UB(src_top); + src_above2 = LD_UB(src_top + 16); + + HADD_UB2_UH(src_above1, src_above2, sum_above1, sum_above2); + + sum_above = sum_above1 + sum_above2; + sum_top = __msa_hadd_u_w(sum_above, sum_above); + sum = __msa_hadd_u_d(sum_top, sum_top); + sum_top = (v4u32) __msa_pckev_w((v4i32) sum, (v4i32) sum); + sum = __msa_hadd_u_d(sum_top, sum_top); + sum = (v2u64) __msa_srari_d((v2i64) sum, 5); + store = (v16u8) __msa_splati_b((v16i8) sum, 0); + } else { + store = (v16u8) __msa_ldi_b(128); + } + for (row = 32; row--;) { + ST_UB2(store, store, dst, 16); + dst += dst_stride; + } +} + +#define INTRA_PREDICT_VALDC_4X4_MSA(val) \ +static void intra_predict_##val##dc_4x4_msa(uint8_t *dst, \ + int32_t dst_stride) \ +{ \ + uint32_t row, out; \ + v16i8 store; \ + \ + store = __msa_ldi_b(val); \ + out = __msa_copy_u_w((v4i32) store, 0); \ + \ + for (row = 4; row--;) \ + { \ + SW(out, dst); \ + dst += dst_stride; \ + } \ +} + +INTRA_PREDICT_VALDC_4X4_MSA(127); +INTRA_PREDICT_VALDC_4X4_MSA(129); + +#define INTRA_PREDICT_VALDC_8X8_MSA(val) \ +static void intra_predict_##val##dc_8x8_msa(uint8_t *dst, \ + int32_t dst_stride) \ +{ \ + uint32_t row, out; \ + v16i8 store; \ + \ + store = __msa_ldi_b(val); \ + out = __msa_copy_u_w((v4i32) store, 0); \ + \ + for (row = 8; row--;) \ + { \ + SW(out, dst); \ + SW(out, (dst + 4)); \ + dst += dst_stride; \ + } \ +} + +INTRA_PREDICT_VALDC_8X8_MSA(127); +INTRA_PREDICT_VALDC_8X8_MSA(129); You've looked too much at corporate h264 decoders (maybe h265 also). FFmpeg gives special indices to the edge dc versions (or the fixed-value dc versions), so that the generic dc version (where both a and l are available) does not need any special branches (your function also isn't marked as inline). I also don't see why you have special versions for dc127/dc129, but don't use that for dc128. Shivraj:-Incorporated changes accordingly. +void ff_vert_4x4_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_vert_4x4_msa(top, dst, stride); +} + +void ff_vert_8x8_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_vert_8x8_msa(top, dst, stride); +} + +void ff_vert_16x16_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_vert_16x16_msa(top, dst, stride); +} + +void ff_vert_32x32_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_vert_32x32_msa(top, dst, stride); +} + +void ff_hor_4x4_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_horiz_4x4_msa(left + 3, -1, dst, stride); +} + +void ff_hor_8x8_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_horiz_8x8_msa(left + 7, -1, dst, stride); +} + +void ff_hor_16x16_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_horiz_16x16_msa(left + 15, -1, dst, stride); +} + +void ff_hor_32x32_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_horiz_32x32_msa(left + 31, -1, dst, stride); +} + +void ff_dc_4x4_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_dc_4x4_msa(top, left, 1, dst, stride, 1, 1); +} + +void ff_dc_8x8_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_dc_8x8_msa(top, left, 1, dst, stride, 1, 1); +} + +void ff_dc_16x16_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_dc_16x16_msa(top, left, 1, dst, stride, 1, 1); +} + +void ff_dc_32x32_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_dc_32x32_msa(top, left, 1, dst, stride, 1, 1); +} + +void ff_dc_left_4x4_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_dc_4x4_msa(top, left, 1, dst, stride, 0, 1); +} + +void ff_dc_left_8x8_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_dc_8x8_msa(top, left, 1, dst, stride, 0, 1); +} + +void ff_dc_left_16x16_msa(uint8_t *dst, ptrdiff_t stride, + const uint8_t *left, const uint8_t *top) +{ + intra_predict_dc_16x16_msa(top, left, 1, dst, stride, 0, 1); +} + +void ff_dc_left_32x32_msa(uint8_t *dst, ptrdiff_t stride, + const uint8_t *left, const uint8_t *top) +{ + intra_predict_dc_32x32_msa(top, left, 1, dst, stride, 0, 1); +} + +void ff_dc_top_4x4_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_dc_4x4_msa(top, left, 1, dst, stride, 1, 0); +} + +void ff_dc_top_8x8_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_dc_8x8_msa(top, left, 1, dst, stride, 1, 0); +} + +void ff_dc_top_16x16_msa(uint8_t *dst, ptrdiff_t stride, + const uint8_t *left, const uint8_t *top) +{ + intra_predict_dc_16x16_msa(top, left, 1, dst, stride, 1, 0); +} + +void ff_dc_top_32x32_msa(uint8_t *dst, ptrdiff_t stride, + const uint8_t *left, const uint8_t *top) +{ + intra_predict_dc_32x32_msa(top, left, 1, dst, stride, 1, 0); +} + +void ff_dc_128_4x4_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_dc_4x4_msa(top, left, 1, dst, stride, 0, 0); +} + +void ff_dc_128_8x8_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_dc_8x8_msa(top, left, 1, dst, stride, 0, 0); +} + +void ff_dc_128_16x16_msa(uint8_t *dst, ptrdiff_t stride, + const uint8_t *left, const uint8_t *top) +{ + intra_predict_dc_16x16_msa(top, left, 1, dst, stride, 0, 0); +} + +void ff_dc_128_32x32_msa(uint8_t *dst, ptrdiff_t stride, + const uint8_t *left, const uint8_t *top) +{ + intra_predict_dc_32x32_msa(top, left, 1, dst, stride, 0, 0); +} + +void ff_dc_127_4x4_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_127dc_4x4_msa(dst, stride); +} + +void ff_dc_127_8x8_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_127dc_8x8_msa(dst, stride); +} + +void ff_dc_127_16x16_msa(uint8_t *dst, ptrdiff_t stride, + const uint8_t *left, const uint8_t *top) +{ + intra_predict_127dc_16x16_msa(dst, stride); +} + +void ff_dc_127_32x32_msa(uint8_t *dst, ptrdiff_t stride, + const uint8_t *left, const uint8_t *top) +{ + intra_predict_127dc_32x32_msa(dst, stride); +} + +void ff_dc_129_4x4_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_129dc_4x4_msa(dst, stride); +} + +void ff_dc_129_8x8_msa(uint8_t *dst, ptrdiff_t stride, const uint8_t *left, + const uint8_t *top) +{ + intra_predict_129dc_8x8_msa(dst, stride); +} + +void ff_dc_129_16x16_msa(uint8_t *dst, ptrdiff_t stride, + const uint8_t *left, const uint8_t *top) +{ + intra_predict_129dc_16x16_msa(dst, stride); +} + +void ff_dc_129_32x32_msa(uint8_t *dst, ptrdiff_t stride, + const uint8_t *left, const uint8_t *top) +{ + intra_predict_129dc_32x32_msa(dst, stride); +} + +void ff_tm_4x4_msa(uint8_t *dst, ptrdiff_t stride, + const uint8_t *left, const uint8_t *top) +{ + intra_predict_tm_4x4_msa(top, left + 3, -1, dst, stride); +} + +void ff_tm_8x8_msa(uint8_t *dst, ptrdiff_t stride, + const uint8_t *left, const uint8_t *top) +{ + intra_predict_tm_8x8_msa(top, left + 7, -1, dst, stride); +} + +void ff_tm_16x16_msa(uint8_t *dst, ptrdiff_t stride, + const uint8_t *left, const uint8_t *top) +{ + intra_predict_tm_16x16_msa(top, left + 15, -1, dst, stride); +} + +void ff_tm_32x32_msa(uint8_t *dst, ptrdiff_t stride, + const uint8_t *left, const uint8_t *top) +{ + intra_predict_tm_32x32_msa(top, left + 31, -1, dst, stride); +} Why do you need these wrappers? Shivraj:- Removed wrappers. Ronald _______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
