anv: add nir lowering pass for ycbcr textures

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.

v2: Add new file to meson build (Lionel)
    Use nir_frcp() rather than (1.0f / x) (Jason)
    Reuse nir_tex_instr_dest_size() rather than handwritten one (Jason)
    Return progress (Jason)
    Account for array of samplers (Jason)

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>

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 \

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);
 
+bool 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,

src/intel/vulkan/anv_nir_lower_ycbcr_textures.c

@@ -0,0 +1,475 @@
+/*
+ * 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_frcp(b, nir_imm_float(b, 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_frcp(b, nir_imm_float(b, 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));
+}
+
+/* TODO: we should probably replace this with a push constant/uniform. */
+static nir_ssa_def *
+get_texture_size(struct ycbcr_state *state, nir_deref_var *texture)
+{
+   if (state->image_size)
+      return state->image_size;
+
+   nir_builder *b = state->builder;
+   const struct glsl_type *type = nir_deref_tail(&texture->deref)->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,
+                     nir_tex_instr_dest_size(tex), 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),
+                            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 = get_texture_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++)
+      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) {
+            assert(old_tex->src[i].src.is_ssa);
+            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 bool
+try_lower_tex_ycbcr(struct anv_pipeline *pipeline,
+                    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];
+
+   /* For the following instructions, we don't apply any change and let the
+    * instruction apply to the first plane.
+    */
+   if (tex->op == nir_texop_txs ||
+       tex->op == nir_texop_query_levels ||
+       tex->op == nir_texop_lod)
+      return false;
+
+   if (binding->immutable_samplers == NULL)
+      return false;
+
+   unsigned texture_index = tex->texture_index;
+   if (tex->texture->deref.child) {
+      assert(tex->texture->deref.child->deref_type == nir_deref_type_array);
+      nir_deref_array *deref_array = nir_deref_as_array(tex->texture->deref.child);
+      if (deref_array->deref_array_type != nir_deref_array_type_direct)
+         return false;
+      size_t hw_binding_size =
+         anv_descriptor_set_binding_layout_get_hw_size(binding);
+      texture_index += MIN2(deref_array->base_offset, hw_binding_size - 1);
+   }
+   const struct anv_sampler *sampler =
+      binding->immutable_samplers[texture_index];
+
+   if (sampler->conversion == NULL)
+      return false;
+
+   struct ycbcr_state state = {
+      .builder = builder,
+      .origin_tex = tex,
+      .conversion = sampler->conversion,
+   };
+
+   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);
+
+   return true;
+}
+
+bool
+anv_nir_lower_ycbcr_textures(nir_shader *shader, struct anv_pipeline *pipeline)
+{
+   bool progress = false;
+
+   nir_foreach_function(function, shader) {
+      if (!function->impl)
+         continue;
+
+      bool function_progress = false;
+      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);
+            function_progress |= try_lower_tex_ycbcr(pipeline, &builder, tex);
+         }
+      }
+
+      if (function_progress) {
+         nir_metadata_preserve(function->impl,
+                               nir_metadata_block_index |
+                               nir_metadata_dominance);
+      }
+
+      progress |= function_progress;
+   }
+
+   return progress;
+}

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)

src/intel/vulkan/anv_private.h

@@ -2544,9 +2544,21 @@ 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 {

src/intel/vulkan/meson.build

@@ -101,6 +101,7 @@ libanv_files = files(
   'anv_nir_lower_input_attachments.c',
   'anv_nir_lower_multiview.c',
   'anv_nir_lower_push_constants.c',
+  'anv_nir_lower_ycbcr_textures.c',
   'anv_pass.c',
   'anv_pipeline.c',
   'anv_pipeline_cache.c',