scaling: Avoid SSSE3 convolution instructions when scaling The SSE3 convolve functions do not work when scaling is involved. If either x_step_q4 or y_step_q4 is not 16 scaling will occur and an assert fails. This patch creates a C version of av1_convolve, av1_convolve_c. This function in turn calls all C versions of what av1_convolve calls. New C versions of these called functions have been created where needed and the same for the functions they call. This means that when scaling is enabled no asserts fail. av1_convolve_c is called instead of av1_convolve when x_step_q4 or y_step_q4 is not 16. Change-Id: I604d2716e034e23a0553fb7004133d3075514a7a
diff --git a/av1/common/convolve.c b/av1/common/convolve.c index cf2acd0..1a8ef88 100644 --- a/av1/common/convolve.c +++ b/av1/common/convolve.c
@@ -131,6 +131,27 @@ } } +void av1_convolve_horiz_facade_c(const uint8_t *src, int src_stride, + uint8_t *dst, int dst_stride, int w, int h, + const InterpFilterParams filter_params, + const int subpel_x_q4, int x_step_q4, + ConvolveParams *conv_params) { + assert(conv_params->round == CONVOLVE_OPT_ROUND); + if (filter_params.taps == SUBPEL_TAPS) { + const int16_t *filter_x = + av1_get_interp_filter_subpel_kernel(filter_params, subpel_x_q4); + if (conv_params->ref == 0) + aom_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter_x, + x_step_q4, NULL, -1, w, h); + else + aom_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride, filter_x, + x_step_q4, NULL, -1, w, h); + } else { + av1_convolve_horiz_c(src, src_stride, dst, dst_stride, w, h, filter_params, + subpel_x_q4, x_step_q4, conv_params); + } +} + void av1_convolve_vert_facade(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams filter_params, @@ -153,6 +174,28 @@ } } +void av1_convolve_vert_facade_c(const uint8_t *src, int src_stride, + uint8_t *dst, int dst_stride, int w, int h, + const InterpFilterParams filter_params, + const int subpel_y_q4, int y_step_q4, + ConvolveParams *conv_params) { + assert(conv_params->round == CONVOLVE_OPT_ROUND); + if (filter_params.taps == SUBPEL_TAPS) { + const int16_t *filter_y = + av1_get_interp_filter_subpel_kernel(filter_params, subpel_y_q4); + if (conv_params->ref == 0) { + aom_convolve8_vert_c(src, src_stride, dst, dst_stride, NULL, -1, filter_y, + y_step_q4, w, h); + } else { + aom_convolve8_avg_vert_c(src, src_stride, dst, dst_stride, NULL, -1, + filter_y, y_step_q4, w, h); + } + } else { + av1_convolve_vert_c(src, src_stride, dst, dst_stride, w, h, filter_params, + subpel_y_q4, y_step_q4, conv_params); + } +} + #if CONFIG_CONVOLVE_ROUND void av1_convolve_rounding(const int32_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, int bits) { @@ -411,6 +454,133 @@ } } +void av1_convolve_c(const uint8_t *src, int src_stride, uint8_t *dst, + int dst_stride, int w, int h, +#if CONFIG_DUAL_FILTER + const InterpFilter *interp_filter, +#else + const InterpFilter interp_filter, +#endif + const int subpel_x_q4, int x_step_q4, const int subpel_y_q4, + int y_step_q4, ConvolveParams *conv_params) { + int ignore_horiz = x_step_q4 == 16 && subpel_x_q4 == 0; + int ignore_vert = y_step_q4 == 16 && subpel_y_q4 == 0; +#if CONFIG_DUAL_FILTER + InterpFilterParams filter_params_x = + av1_get_interp_filter_params(interp_filter[1 + 2 * conv_params->ref]); + InterpFilterParams filter_params_y = + av1_get_interp_filter_params(interp_filter[0 + 2 * conv_params->ref]); + InterpFilterParams filter_params; +#else + InterpFilterParams filter_params = + av1_get_interp_filter_params(interp_filter); +#endif + assert(conv_params->round == CONVOLVE_OPT_ROUND); + + assert(w <= MAX_BLOCK_WIDTH); + assert(h <= MAX_BLOCK_HEIGHT); + assert(y_step_q4 <= MAX_STEP); + assert(x_step_q4 <= MAX_STEP); + + if (ignore_horiz && ignore_vert) { + convolve_copy(src, src_stride, dst, dst_stride, w, h, conv_params); + } else if (ignore_vert) { +#if CONFIG_DUAL_FILTER + filter_params = filter_params_x; +#endif + assert(filter_params.taps <= MAX_FILTER_TAP); + av1_convolve_horiz_facade_c(src, src_stride, dst, dst_stride, w, h, + filter_params, subpel_x_q4, x_step_q4, + conv_params); + } else if (ignore_horiz) { +#if CONFIG_DUAL_FILTER + filter_params = filter_params_y; +#endif + assert(filter_params.taps <= MAX_FILTER_TAP); + av1_convolve_vert_facade_c(src, src_stride, dst, dst_stride, w, h, + filter_params, subpel_y_q4, y_step_q4, + conv_params); + } else { + // temp's size is set to a 256 aligned value to facilitate SIMD + // implementation. The value is greater than (maximum possible intermediate + // height or width) * MAX_SB_SIZE + DECLARE_ALIGNED(16, uint8_t, + temp[((MAX_SB_SIZE * 2 + 16) + 16) * MAX_SB_SIZE]); + int max_intermediate_size = ((MAX_SB_SIZE * 2 + 16) + 16); + int filter_size; +#if CONFIG_DUAL_FILTER + if (interp_filter[0 + 2 * conv_params->ref] == MULTITAP_SHARP && + interp_filter[1 + 2 * conv_params->ref] == MULTITAP_SHARP) { + // Avoid two directions both using 12-tap filter. + // This will reduce hardware implementation cost. + filter_params_y = av1_get_interp_filter_params(EIGHTTAP_SHARP); + } + + // we do filter with fewer taps first to reduce hardware implementation + // complexity + if (filter_params_y.taps < filter_params_x.taps) { + int intermediate_width; + int temp_stride = max_intermediate_size; + ConvolveParams temp_conv_params; + temp_conv_params.ref = 0; + temp_conv_params.round = CONVOLVE_OPT_ROUND; + filter_params = filter_params_y; + filter_size = filter_params_x.taps; + intermediate_width = + (((w - 1) * x_step_q4 + subpel_x_q4) >> SUBPEL_BITS) + filter_size; + assert(intermediate_width <= max_intermediate_size); + + assert(filter_params.taps <= MAX_FILTER_TAP); + + av1_convolve_vert_facade(src - (filter_size / 2 - 1), src_stride, temp, + temp_stride, intermediate_width, h, + filter_params, subpel_y_q4, y_step_q4, + &temp_conv_params); + + filter_params = filter_params_x; + assert(filter_params.taps <= MAX_FILTER_TAP); + av1_convolve_horiz_facade(temp + (filter_size / 2 - 1), temp_stride, dst, + dst_stride, w, h, filter_params, subpel_x_q4, + x_step_q4, conv_params); + } else { +#endif // CONFIG_DUAL_FILTER + int intermediate_height; + int temp_stride = MAX_SB_SIZE; + ConvolveParams temp_conv_params; + temp_conv_params.ref = 0; + temp_conv_params.round = CONVOLVE_OPT_ROUND; +#if CONFIG_DUAL_FILTER + filter_params = filter_params_x; + filter_size = filter_params_y.taps; +#else + filter_size = filter_params.taps; +#endif + intermediate_height = + (((h - 1) * y_step_q4 + subpel_y_q4) >> SUBPEL_BITS) + filter_size; + assert(intermediate_height <= max_intermediate_size); + (void)max_intermediate_size; + + assert(filter_params.taps <= MAX_FILTER_TAP); + + av1_convolve_horiz_facade_c(src - src_stride * (filter_size / 2 - 1), + src_stride, temp, temp_stride, w, + intermediate_height, filter_params, + subpel_x_q4, x_step_q4, &temp_conv_params); + +#if CONFIG_DUAL_FILTER + filter_params = filter_params_y; +#endif + assert(filter_params.taps <= MAX_FILTER_TAP); + + av1_convolve_vert_facade_c( + temp + temp_stride * (filter_size / 2 - 1), temp_stride, dst, + dst_stride, w, h, filter_params, subpel_y_q4, y_step_q4, conv_params); +#if CONFIG_DUAL_FILTER + } +#endif // CONFIG_DUAL_FILTER + } +} + void av1_lowbd_convolve_init_c(void) { // A placeholder for SIMD initialization return;
diff --git a/av1/common/convolve.h b/av1/common/convolve.h index 471152e..cb6569c 100644 --- a/av1/common/convolve.h +++ b/av1/common/convolve.h
@@ -86,6 +86,16 @@ const int subpel_x, int xstep, const int subpel_y, int ystep, ConvolveParams *conv_params); +void av1_convolve_c(const uint8_t *src, int src_stride, uint8_t *dst, + int dst_stride, int w, int h, +#if CONFIG_DUAL_FILTER + const InterpFilter *interp_filter, +#else + const InterpFilter interp_filter, +#endif + const int subpel_x, int xstep, const int subpel_y, + int ystep, ConvolveParams *conv_params); + #if CONFIG_AOM_HIGHBITDEPTH void av1_highbd_convolve(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h,
diff --git a/av1/common/reconinter.h b/av1/common/reconinter.h index 5408870..c33c936 100644 --- a/av1/common/reconinter.h +++ b/av1/common/reconinter.h
@@ -56,14 +56,14 @@ #if CONFIG_DUAL_FILTER if (interp_filter_params_x.taps == SUBPEL_TAPS && interp_filter_params_y.taps == SUBPEL_TAPS && w > 2 && h > 2 && - conv_params->round == CONVOLVE_OPT_ROUND) { + conv_params->round == CONVOLVE_OPT_ROUND && xs == 16 && ys == 16) { const int16_t *kernel_x = av1_get_interp_filter_subpel_kernel(interp_filter_params_x, subpel_x); const int16_t *kernel_y = av1_get_interp_filter_subpel_kernel(interp_filter_params_y, subpel_y); #else if (interp_filter_params.taps == SUBPEL_TAPS && w > 2 && h > 2 && - conv_params->round == CONVOLVE_OPT_ROUND) { + conv_params->round == CONVOLVE_OPT_ROUND && xs == 16 && ys == 16) { const int16_t *kernel_x = av1_get_interp_filter_subpel_kernel(interp_filter_params, subpel_x); const int16_t *kernel_y = @@ -76,7 +76,7 @@ // first reference frame's prediction result is already in dst // therefore we need to average the first and second results #if CONFIG_CONVOLVE_ROUND - if (conv_params->round == CONVOLVE_OPT_NO_ROUND) + if (conv_params->round == CONVOLVE_OPT_NO_ROUND && xs == 16 && ys == 16) av1_convolve_2d_facade(src, src_stride, dst, dst_stride, w, h, #if CONFIG_DUAL_FILTER interp_filter, @@ -86,8 +86,17 @@ subpel_x, xs, subpel_y, ys, conv_params); else #endif - av1_convolve(src, src_stride, dst, dst_stride, w, h, interp_filter, - subpel_x, xs, subpel_y, ys, conv_params); + { + if (xs == 16 && ys == 16) { + av1_convolve(src, src_stride, dst, dst_stride, w, h, interp_filter, + subpel_x, xs, subpel_y, ys, conv_params); + } else { + // If xs == 16 || ys == 16 scaling is happening and the SSE2 + // instructions don't support scaling; use the C versions to be safe. + av1_convolve_c(src, src_stride, dst, dst_stride, w, h, interp_filter, + subpel_x, xs, subpel_y, ys, conv_params); + } + } } }