<div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Tue, Oct 3, 2017 at 4:54 PM, Jason Ekstrand <<a href="mailto:jason@jlekstrand.net" target="_blank">jason@jlekstrand.net</a>> wrote: <blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div><div class="h5">On Tue, Oct 3, 2017 at 9:29 AM, Lionel Landwerlin <<a href="mailto:lionel.g.landwerlin@intel.com" target="_blank">lionel.g.landwerlin@intel.com</a>> wrote: <blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">This pass implements all the implicit conversions required by the VK_KHR_sampler_ycbcr_conversion specification. It also inserts plane sources onto sampling instructions that we then let the pipeline layout pass deal with, when mapping things correctly to descriptors. Signed-off-by: Lionel Landwerlin <<a href="mailto:lionel.g.landwerlin@intel.com" target="_blank">lionel.g.landwerlin@intel.com</a>> --- src/intel/Makefile.sources | 1 + src/intel/vulkan/anv_nir.h | 3 + src/intel/vulkan/anv_nir_apply_pipeline_layout.c | 62 ++- src/intel/vulkan/anv_nir_lower_ycbcr_textures.c | 468 +++++++++++++++++++++++ src/intel/vulkan/anv_pipeline.c | 2 + src/intel/vulkan/anv_private.h | 16 +- 6 files changed, 545 insertions(+), 7 deletions(-) create mode 100644 src/intel/vulkan/anv_nir_lower_ycbcr_textures.c diff --git a/src/intel/Makefile.sources b/src/intel/Makefile.sources index bca7a132b26..9672dcc252d 100644 --- a/src/intel/Makefile.sources +++ b/src/intel/Makefile.sources @@ -219,6 +219,7 @@ VULKAN_FILES := \ vulkan/anv_nir_lower_input_attachments.c \ vulkan/anv_nir_lower_multiview.c \ vulkan/anv_nir_lower_push_constants.c \ + vulkan/anv_nir_lower_ycbcr_textures.c \ vulkan/anv_pass.c \ vulkan/anv_pipeline.c \ vulkan/anv_pipeline_cache.c \ diff --git a/src/intel/vulkan/anv_nir.h b/src/intel/vulkan/anv_nir.h index 5b450b45cdf..0a06e3a1cf0 100644 --- a/src/intel/vulkan/anv_nir.h +++ b/src/intel/vulkan/anv_nir.h @@ -37,6 +37,9 @@ void anv_nir_lower_push_constants(nir_shader *shader); bool anv_nir_lower_multiview(nir_shader *shader, uint32_t view_mask); +void anv_nir_lower_ycbcr_textures(nir_shader *shader, + struct anv_pipeline *pipeline); + void anv_nir_apply_pipeline_layout(struct anv_pipeline *pipeline, nir_shader *shader, struct brw_stage_prog_data *prog_data, diff --git a/src/intel/vulkan/anv_nir_apply_pipeline_layout.c b/src/intel/vulkan/anv_nir_apply_pipeline_layout.c index 428cfdf42d1..7cd28debe09 100644 --- a/src/intel/vulkan/anv_nir_apply_pipeline_layout.c +++ b/src/intel/vulkan/anv_nir_apply_pipeline_layout.c @@ -131,7 +131,7 @@ lower_res_index_intrinsic(nir_intrinsic_instr *intrin, static void lower_tex_deref(nir_tex_instr *tex, nir_deref_var *deref, unsigned *const_index, unsigned hw_binding_size, - nir_tex_src_type src_type, + nir_tex_src_type src_type, bool allow_indirect, struct apply_pipeline_layout_state *state) { nir_builder *b = &state->builder; @@ -141,6 +141,15 @@ lower_tex_deref(nir_tex_instr *tex, nir_deref_var *deref, nir_deref_array *deref_array = nir_deref_as_array(deref->deref.child); if (deref_array->deref_array_type == nir_deref_array_type_indirect) { + /* From VK_KHR_sampler_ycbcr_conversion: + * + * If sampler Y’CBCR conversion is enabled, the combined image + * sampler must be indexed only by constant integral expressions when + * aggregated into arrays in shader code, irrespective of the + * shaderSampledImageArrayDynamicIndexing feature. + */ + assert(allow_indirect); + nir_ssa_def *index = nir_iadd(b, nir_imm_int(b, deref_array->base_offset), nir_ssa_for_src(b, deref_array->indirect, 1)); @@ -150,7 +159,6 @@ lower_tex_deref(nir_tex_instr *tex, nir_deref_var *deref, nir_tex_src *new_srcs = rzalloc_array(tex, nir_tex_src, tex->num_srcs + 1); - </blockquote><div> </div></div></div><div>Spurrious change? </div><div><div class="h5"><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> for (unsigned i = 0; i < tex->num_srcs; i++) { new_srcs[i].src_type = tex->src[i].src_type; nir_instr_move_src(&tex->instr, &new_srcs[i].src, &tex->src[i].src); @@ -186,6 +194,46 @@ cleanup_tex_deref(nir_tex_instr *tex, nir_deref_var *deref) nir_instr_rewrite_src(&tex->instr, &deref_array->indirect, NIR_SRC_INIT); } +static bool +has_tex_src_plane(nir_tex_instr *tex) +{ + for (unsigned i = 0; i < tex->num_srcs; i++) { + if (tex->src[i].src_type == nir_tex_src_plane) + return true; + } + + return false; +} + +static uint32_t +extract_tex_src_plane(nir_tex_instr *tex) +{ + nir_tex_src *new_srcs = rzalloc_array(tex, nir_tex_src, tex->num_srcs - 1); + unsigned plane = 0; + + for (unsigned i = 0, w = 0; i < tex->num_srcs; i++) { + if (tex->src[i].src_type == nir_tex_src_plane) { + nir_const_value *const_plane = + nir_src_as_const_value(tex->src[i].src); + + /* Our color conversion lowering pass should only ever insert + * constants. */ + assert(const_plane); + plane = const_plane->u32[0]; + } else { + new_srcs[w].src_type = tex->src[i].src_type; + nir_instr_move_src(&tex->instr, &new_srcs[w].src, &tex->src[i].src); + w++; + } + } + + ralloc_free(tex->src); + tex->src = new_srcs; + tex->num_srcs--; + + return plane; +} + static void lower_tex(nir_tex_instr *tex, struct apply_pipeline_layout_state *state) { @@ -198,17 +246,21 @@ lower_tex(nir_tex_instr *tex, struct apply_pipeline_layout_state *state) unsigned binding = tex->texture->var->data.binding; unsigned hw_binding_size = state->layout->set[set].layout->binding[binding].hw_binding_size; - tex->texture_index = state->set[set].surface_offsets[binding]; + bool has_plane = has_tex_src_plane(tex); + unsigned plane = has_plane ? extract_tex_src_plane(tex) : 0; + + tex->texture_index = state->set[set].surface_offsets[binding] + plane; lower_tex_deref(tex, tex->texture, &tex->texture_index, hw_binding_size, - nir_tex_src_texture_offset, state); + nir_tex_src_texture_offset, !has_plane, state); if (tex->sampler) { unsigned set = tex->sampler->var->data.descriptor_set; unsigned binding = tex->sampler->var->data.binding; unsigned hw_binding_size = state->layout->set[set].layout->binding[binding].hw_binding_size; + tex->sampler_index = state->set[set].sampler_offsets[binding] + plane; lower_tex_deref(tex, tex->sampler, &tex->sampler_index, hw_binding_size, - nir_tex_src_sampler_offset, state); + nir_tex_src_sampler_offset, !has_plane, state); } /* The backend only ever uses this to mark used surfaces. We don't care diff --git a/src/intel/vulkan/anv_nir_lower_ycbcr_textures.c b/src/intel/vulkan/anv_nir_lower_ycbcr_textures.c new file mode 100644 index 00000000000..cae668cba2e --- /dev/null +++ b/src/intel/vulkan/anv_nir_lower_ycbcr_textures.c @@ -0,0 +1,468 @@ +/* + * Copyright © 2017 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + */ + +#include "anv_nir.h" +#include "anv_private.h" +#include "nir.h" +#include "nir/nir_builder.h" + +struct ycbcr_state { + nir_builder *builder; + nir_ssa_def *image_size; + nir_tex_instr *origin_tex; + struct anv_ycbcr_conversion *conversion; +}; + +static nir_ssa_def * +y_range(nir_builder *b, + nir_ssa_def *y_channel, + int bpc, + VkSamplerYcbcrRangeKHR range) +{ + switch (range) { + case VK_SAMPLER_YCBCR_RANGE_ITU_FULL_KHR: + return y_channel; + case VK_SAMPLER_YCBCR_RANGE_ITU_NARROW_KHR: + return nir_fmul(b, + nir_fadd(b, + nir_fmul(b, y_channel, + nir_imm_float(b, pow(2, bpc) - 1)), + nir_imm_float(b, -16.0f * pow(2, bpc - 8))), + nir_imm_float(b, 1.0f / (219.0f * pow(2, bpc - 8)))); + default: + unreachable("missing Ycbcr range"); + return NULL; + } +} + +static nir_ssa_def * +chroma_range(nir_builder *b, + nir_ssa_def *chroma_channel, + int bpc, + VkSamplerYcbcrRangeKHR range) +{ + switch (range) { + case VK_SAMPLER_YCBCR_RANGE_ITU_FULL_KHR: + return nir_fadd(b, chroma_channel, + nir_imm_float(b, -pow(2, bpc - 1) / (pow(2, bpc) - 1.0f))); + case VK_SAMPLER_YCBCR_RANGE_ITU_NARROW_KHR: + return nir_fmul(b, + nir_fadd(b, + nir_fmul(b, chroma_channel, + nir_imm_float(b, pow(2, bpc) - 1)), + nir_imm_float(b, -128.0f * pow(2, bpc - 8))), + nir_imm_float(b, 1.0f / (224.0f * pow(2, bpc - 8)))); + default: + unreachable("missing Ycbcr range"); + return NULL; + } +} + +static const nir_const_value * +ycbcr_model_to_rgb_matrix(VkSamplerYcbcrModelConversionKHR model) +{ + switch (model) { + case VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_601_KHR: { + static const nir_const_value bt601[3] = { + { .f32 = { 1.402f, 1.0f, 0.0f, 0.0f } }, + { .f32 = { -0.714136286201022f, 1.0f, -0.344136286201022f, 0.0f } }, + { .f32 = { 0.0f, 1.0f, 1.772f, 0.0f } } + }; + + return bt601; + } + case VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_709_KHR: { + static const nir_const_value bt709[3] = { + { .f32 = { 1.5748031496063f, 1.0f, 0.0, 0.0f } }, + { .f32 = { -0.468125209181067f, 1.0f, -0.187327487470334f, 0.0f } }, + { .f32 = { 0.0f, 1.0f, 1.85563184264242f, 0.0f } } + }; + + return bt709; + } + case VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_2020_KHR: { + static const nir_const_value bt2020[3] = { + { .f32 = { 1.4746f, 1.0f, 0.0f, 0.0f } }, + { .f32 = { -0.571353126843658f, 1.0f, -0.164553126843658f, 0.0f } }, + { .f32 = { 0.0f, 1.0f, 1.8814f, 0.0f } } + }; + + return bt2020; + } + default: + unreachable("missing Ycbcr model"); + return NULL; + } +} + +static nir_ssa_def * +convert_ycbcr(struct ycbcr_state *state, + nir_ssa_def *raw_channels, + uint32_t *bpcs) +{ + nir_builder *b = state->builder; + struct anv_ycbcr_conversion *conversion = state->conversion; + + nir_ssa_def *expanded_channels = + nir_vec4(b, + chroma_range(b, nir_channel(b, raw_channels, 0), + bpcs[0], conversion->ycbcr_range), + y_range(b, nir_channel(b, raw_channels, 1), + bpcs[1], conversion->ycbcr_range), + chroma_range(b, nir_channel(b, raw_channels, 2), + bpcs[2], conversion->ycbcr_range), + nir_imm_float(b, 1.0f)); + + if (conversion->ycbcr_model == VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_IDENTITY_KHR) + return expanded_channels; + + const nir_const_value *conversion_matrix = + ycbcr_model_to_rgb_matrix(conversion->ycbcr_model); + + nir_ssa_def *converted_channels[] = { + nir_fdot4(b, expanded_channels, nir_build_imm(b, 4, 32, conversion_matrix[0])), + nir_fdot4(b, expanded_channels, nir_build_imm(b, 4, 32, conversion_matrix[1])), + nir_fdot4(b, expanded_channels, nir_build_imm(b, 4, 32, conversion_matrix[2])) + }; + + return nir_vec4(b, + converted_channels[0], converted_channels[1], + converted_channels[2], nir_imm_float(b, 1.0f)); +} + +static unsigned +sampler_dim_components(enum glsl_sampler_dim sampler_dim) +{ + switch (sampler_dim) { + case GLSL_SAMPLER_DIM_1D: + case GLSL_SAMPLER_DIM_BUF: + return 1; + case GLSL_SAMPLER_DIM_RECT: + case GLSL_SAMPLER_DIM_2D: + return 2; + case GLSL_SAMPLER_DIM_3D: + return 3; + case GLSL_SAMPLER_DIM_CUBE: + return 3; + default: + unreachable("unsupported sampler dimension"); + return 0; + } +} </blockquote><div> </div></div></div><div>You don't need this. Just use nir_tex_instr_dest_size(). </div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> + +/* TODO: we should probably replace this with a push constant/uniform. */ +static nir_ssa_def * +maybe_load_image_size(struct ycbcr_state *state, nir_deref_var *texture) </blockquote><div> </div><div>I don't really like the maybe_ Why not just call it "get_texture_size" or something like that. </div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> +{ + if (state->image_size) + return state->image_size; </blockquote><div> </div><div>Don't worry too much about not re-emitting image_size. NIR will CSE any duplicate txs operations you may have. </div><div><div class="h5"><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> + + nir_builder *b = state->builder; + const struct glsl_type *type = nir_deref_tail(&texture->deref)->type; + const unsigned num_components = + sampler_dim_components(glsl_get_sampler_dim(type)); + nir_tex_instr *tex = nir_tex_instr_create(b->shader, 0); + + tex->op = nir_texop_txs; + tex->sampler_dim = glsl_get_sampler_dim(type); + tex->is_array = glsl_sampler_type_is_array(type); + tex->is_shadow = glsl_sampler_type_is_shadow(type); + tex->texture = nir_deref_var_clone(texture, tex); + tex->dest_type = nir_type_int; + + nir_ssa_dest_init(&tex->instr, &tex->dest, + num_components, 32, NULL); + nir_builder_instr_insert(b, &tex->instr); + + state->image_size = nir_i2f32(b, &tex->dest.ssa); + + return state->image_size; +} + +static nir_ssa_def * +implicit_downsampled_coord(nir_builder *b, + nir_ssa_def *value, + nir_ssa_def *max_value, + int div_scale) +{ + return nir_fadd(b, + value, + nir_fdiv(b, + nir_imm_float(b, 1.0f), </blockquote><div> </div></div></div><div>nir_frcp </div><div><div class="h5"><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> + nir_fmul(b, + nir_imm_float(b, div_scale), + max_value))); +} + +static nir_ssa_def * +implicit_downsampled_coords(struct ycbcr_state *state, + nir_ssa_def *old_coords, + const struct anv_format_plane *plane_format) +{ + nir_builder *b = state->builder; + struct anv_ycbcr_conversion *conversion = state->conversion; + nir_ssa_def *image_size = maybe_load_image_size(state, + state->origin_tex->texture); + nir_ssa_def *comp[4] = { NULL, }; + int c; + + for (c = 0; c < ARRAY_SIZE(conversion->chroma_offsets); c++) { + if (plane_format->denominator_scales[c] > 1 && + conversion->chroma_offsets[c] == VK_CHROMA_LOCATION_COSITED_EVEN_KHR) { + comp[c] = implicit_downsampled_coord(b, + nir_channel(b, old_coords, c), + nir_channel(b, image_size, c), + plane_format->denominator_scales[c]); + } else { + comp[c] = nir_channel(b, old_coords, c); + } + } + + /* Leave other coordinates untouched */ + for (; c < old_coords->num_components; c++) </blockquote><div> </div></div></div><div>Is non-2D YCBCR allowed? I guess it may make sense to allow arrays.</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> + comp[c] = nir_channel(b, old_coords, c); + + return nir_vec(b, comp, old_coords->num_components); +} + +static nir_ssa_def * +create_plane_tex_instr_implicit(struct ycbcr_state *state, + uint32_t plane) +{ + nir_builder *b = state->builder; + struct anv_ycbcr_conversion *conversion = state->conversion; + const struct anv_format_plane *plane_format = + &conversion->format->planes[plane]; + nir_tex_instr *old_tex = state->origin_tex; + nir_tex_instr *tex = nir_tex_instr_create(b->shader, old_tex->num_srcs + 1); + + for (uint32_t i = 0; i < old_tex->num_srcs; i++) { + tex->src[i].src_type = old_tex->src[i].src_type; + + switch (old_tex->src[i].src_type) { + case nir_tex_src_coord: + if (plane_format->has_chroma && conversion->chroma_reconstruction) { </blockquote><div> </div><div>assert(old_tex->src[i].src.is_ssa); </div><div><div class="h5"><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> + tex->src[i].src = + nir_src_for_ssa(implicit_downsampled_coords(state, + old_tex->src[i].src.ssa, + plane_format)); + break; + } + /* fall through */ + default: + nir_src_copy(&tex->src[i].src, &old_tex->src[i].src, tex); + break; + } + } + tex->src[tex->num_srcs - 1].src = nir_src_for_ssa(nir_imm_int(b, plane)); + tex->src[tex->num_srcs - 1].src_type = nir_tex_src_plane; + + tex->sampler_dim = old_tex->sampler_dim; + tex->dest_type = old_tex->dest_type; + + tex->op = old_tex->op; + tex->coord_components = old_tex->coord_components; + tex->is_new_style_shadow = old_tex->is_new_style_shadow; + tex->component = old_tex->component; + + tex->texture_index = old_tex->texture_index; + tex->texture_array_size = old_tex->texture_array_size; + tex->texture = nir_deref_var_clone(old_tex->texture, tex); + + tex->sampler_index = old_tex->sampler_index; + tex->sampler = nir_deref_var_clone(old_tex->sampler, tex); + + nir_ssa_dest_init(&tex->instr, &tex->dest, + old_tex->dest.ssa.num_components, + nir_dest_bit_size(old_tex->dest), NULL); + nir_builder_instr_insert(b, &tex->instr); + + return &tex->dest.ssa; +} + +static unsigned +channel_to_component(enum isl_channel_select channel) +{ + switch (channel) { + case ISL_CHANNEL_SELECT_RED: + return 0; + case ISL_CHANNEL_SELECT_GREEN: + return 1; + case ISL_CHANNEL_SELECT_BLUE: + return 2; + case ISL_CHANNEL_SELECT_ALPHA: + return 3; + default: + unreachable("invalid channel"); + return 0; + } +} + +static enum isl_channel_select +swizzle_channel(struct isl_swizzle swizzle, unsigned channel) +{ + switch (channel) { + case 0: + return swizzle.r; + case 1: + return swizzle.g; + case 2: + return swizzle.b; + case 3: + return swizzle.a; + default: + unreachable("invalid channel"); + return 0; + } +} + +static void +maybe_replace_texture_multiplanar(struct anv_pipeline *pipeline, </blockquote><div> </div></div></div><div>How about try_lower_tex_ycbcr? </div><div><div class="h5"><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> + nir_builder *builder, + nir_tex_instr *tex) +{ + nir_variable *var = tex->texture->var; + const struct anv_descriptor_set_layout *set_layout = + pipeline->layout->set[var->data.descriptor_set].layout; + const struct anv_descriptor_set_binding_layout *binding = + &set_layout->binding[var->data.binding]; + + if (tex->op == nir_texop_txs || + tex->op == nir_texop_query_levels || + tex->op == nir_texop_lod) + return; + + if (binding->immutable_samplers == NULL) + return; + + const struct anv_sampler *sampler = + binding->immutable_samplers[tex->texture_index]; + + if (sampler->conversion == NULL) + return; + + struct ycbcr_state state = { + .builder = builder, + .origin_tex = tex, + .conversion = sampler->conversion, </blockquote></div></div></div></div></div></blockquote><div> </div><div>What about arrays of samplers? You're not allowed to indirect on them but they can, in theory, exist and have different samplers with different conversions. I think we need to fish the array index out and take sampler[idx].conversion here instead. </div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div><div class="h5"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> + }; + + builder->cursor = nir_before_instr(&tex->instr); + + const struct anv_format *format = state.conversion->format; + const struct isl_format_layout *y_isl_layout = NULL; + for (uint32_t p = 0; p < format->n_planes; p++) { + if (!format->planes[p].has_chroma) + y_isl_layout = isl_format_get_layout(format->planes[p].isl_format); + } + assert(y_isl_layout != NULL); + uint8_t y_bpc = y_isl_layout->channels_array[0].bits; + + /* |ycbcr_comp| holds components in the order : Cr-Y-Cb */ + nir_ssa_def *ycbcr_comp[5] = { NULL, NULL, NULL, + /* Use extra 2 channels for following swizzle */ + nir_imm_float(builder, 1.0f), + nir_imm_float(builder, 0.0f), + }; + uint8_t ycbcr_bpcs[5]; + memset(ycbcr_bpcs, y_bpc, sizeof(ycbcr_bpcs)); + + /* Go through all the planes and gather the samples into a |ycbcr_comp| + * while applying a swizzle required by the spec: + * + * R, G, B should respectively map to Cr, Y, Cb + */ + for (uint32_t p = 0; p < format->n_planes; p++) { + const struct anv_format_plane *plane_format = &format->planes[p]; + nir_ssa_def *plane_sample = create_plane_tex_instr_implicit(&state, p); + + for (uint32_t pc = 0; pc < 4; pc++) { + enum isl_channel_select ycbcr_swizzle = + swizzle_channel(plane_format->ycbcr_swizzle, pc); + if (ycbcr_swizzle == ISL_CHANNEL_SELECT_ZERO) + continue; + + unsigned ycbcr_component = channel_to_component(ycbcr_swizzle); + ycbcr_comp[ycbcr_component] = nir_channel(builder, plane_sample, pc); + + /* Also compute the number of bits for each component. */ + const struct isl_format_layout *isl_layout = + isl_format_get_layout(plane_format->isl_format); + ycbcr_bpcs[ycbcr_component] = isl_layout->channels_array[pc].bits; + } + } + + /* Now remaps components to the order specified by the conversion. */ + nir_ssa_def *swizzled_comp[4] = { NULL, }; + uint32_t swizzled_bpcs[4] = { 0, }; + + for (uint32_t i = 0; i < ARRAY_SIZE(state.conversion->mapping); i++) { + /* Maps to components in |ycbcr_comp| */ + static const uint32_t swizzle_mapping[] = { + [VK_COMPONENT_SWIZZLE_ZERO] = 4, + [VK_COMPONENT_SWIZZLE_ONE] = 3, + [VK_COMPONENT_SWIZZLE_R] = 0, + [VK_COMPONENT_SWIZZLE_G] = 1, + [VK_COMPONENT_SWIZZLE_B] = 2, + [VK_COMPONENT_SWIZZLE_A] = 3, + }; + const VkComponentSwizzle m = state.conversion->mapping[i]; + + if (m == VK_COMPONENT_SWIZZLE_IDENTITY) { + swizzled_comp[i] = ycbcr_comp[i]; + swizzled_bpcs[i] = ycbcr_bpcs[i]; + } else { + swizzled_comp[i] = ycbcr_comp[swizzle_mapping[m]]; + swizzled_bpcs[i] = ycbcr_bpcs[swizzle_mapping[m]]; + } + } + + nir_ssa_def *result = nir_vec(builder, swizzled_comp, 4); + if (state.conversion->ycbcr_model != VK_SAMPLER_YCBCR_MODEL_CONVERSION_RGB_IDENTITY_KHR) + result = convert_ycbcr(&state, result, swizzled_bpcs); + + nir_ssa_def_rewrite_uses(&tex->dest.ssa, nir_src_for_ssa(result)); + nir_instr_remove(&tex->instr); +} + +void +anv_nir_lower_ycbcr_textures(nir_shader *shader, struct anv_pipeline *pipeline) </blockquote><div> </div></div></div><div>Most NIR lowering passes return a boolean that indicates whether or not they actually changed anything. We should probably do the same here. It shouldn't be a big change. </div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> +{ + nir_foreach_function(function, shader) { + if (!function->impl) + continue; + + nir_builder builder; + nir_builder_init(&builder, function->impl); + + nir_foreach_block(block, function->impl) { + nir_foreach_instr_safe(instr, block) { + if (instr->type != nir_instr_type_tex) + continue; + + nir_tex_instr *tex = nir_instr_as_tex(instr); + maybe_replace_texture_multiplanar(pipeline, &builder, tex); + } + } </blockquote><div> </div><div>We need to call nir_metadata_preserve here. </div><div><div class="h5"><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> + } +} diff --git a/src/intel/vulkan/anv_pipeline.c b/src/intel/vulkan/anv_pipeline.c index 94e99d8437a..b35bad10503 100644 --- a/src/intel/vulkan/anv_pipeline.c +++ b/src/intel/vulkan/anv_pipeline.c @@ -376,6 +376,8 @@ anv_pipeline_compile(struct anv_pipeline *pipeline, if (nir == NULL) return NULL; + NIR_PASS_V(nir, anv_nir_lower_ycbcr_textures, pipeline); + NIR_PASS_V(nir, anv_nir_lower_push_constants); if (stage != MESA_SHADER_COMPUTE) diff --git a/src/intel/vulkan/anv_private.h b/src/intel/vulkan/anv_private.h index 63b803df6ea..5e14f49b196 100644 --- a/src/intel/vulkan/anv_private.h +++ b/src/intel/vulkan/anv_private.h @@ -2567,9 +2567,21 @@ void anv_fill_buffer_surface_state(struct anv_device *device, uint32_t offset, uint32_t range, uint32_t stride); + +struct anv_ycbcr_conversion { + const struct anv_format * format; + VkSamplerYcbcrModelConversionKHR ycbcr_model; + VkSamplerYcbcrRangeKHR ycbcr_range; + VkComponentSwizzle mapping[4]; + VkChromaLocationKHR chroma_offsets[2]; + VkFilter chroma_filter; + bool chroma_reconstruction; +}; + struct anv_sampler { - uint32_t state[3][4]; - uint32_t n_planes; + uint32_t state[3][4]; + uint32_t n_planes; + struct anv_ycbcr_conversion *conversion; }; struct anv_framebuffer { -- 2.14.2 _______________________________________________ mesa-dev mailing list <a href="mailto:mesa-dev@lists.freedesktop.org" target="_blank">mesa-dev@lists.freedesktop.org</a> <a href="https://lists.freedesktop.org/mailman/listinfo/mesa-dev" rel="noreferrer" target="_blank">https://lists.freedesktop.org/mailman/listinfo/mesa-dev</a> </blockquote></div></div></div> </div></div> </blockquote></div> </div></div>