anv/pipeline: Call nir_gather_info later
[mesa.git] / src / intel / vulkan / anv_pipeline.c
index bdc2f01351f25d6ea70206c05b4f200ee34d1931..599286c216945f1783f4a1464e5cb939c46bfb00 100644 (file)
@@ -30,7 +30,7 @@
 #include "util/mesa-sha1.h"
 #include "common/gen_l3_config.h"
 #include "anv_private.h"
-#include "brw_nir.h"
+#include "compiler/brw_nir.h"
 #include "anv_nir.h"
 #include "spirv/nir_spirv.h"
 
@@ -75,6 +75,9 @@ void anv_DestroyShaderModule(
    ANV_FROM_HANDLE(anv_device, device, _device);
    ANV_FROM_HANDLE(anv_shader_module, module, _module);
 
+   if (!module)
+      return;
+
    vk_free2(&device->alloc, pAllocator, module);
 }
 
@@ -84,15 +87,13 @@ void anv_DestroyShaderModule(
  * we can't do that yet because we don't have the ability to copy nir.
  */
 static nir_shader *
-anv_shader_compile_to_nir(struct anv_device *device,
+anv_shader_compile_to_nir(struct anv_pipeline *pipeline,
                           struct anv_shader_module *module,
                           const char *entrypoint_name,
                           gl_shader_stage stage,
                           const VkSpecializationInfo *spec_info)
 {
-   if (strcmp(entrypoint_name, "main") != 0) {
-      anv_finishme("Multiple shaders per module not really supported");
-   }
+   const struct anv_device *device = pipeline->device;
 
    const struct brw_compiler *compiler =
       device->instance->physicalDevice.compiler;
@@ -114,30 +115,38 @@ anv_shader_compile_to_nir(struct anv_device *device,
          assert(data + entry.size <= spec_info->pData + spec_info->dataSize);
 
          spec_entries[i].id = spec_info->pMapEntries[i].constantID;
-         spec_entries[i].data = *(const uint32_t *)data;
+         if (spec_info->dataSize == 8)
+            spec_entries[i].data64 = *(const uint64_t *)data;
+         else
+            spec_entries[i].data32 = *(const uint32_t *)data;
       }
    }
 
+   const struct nir_spirv_supported_extensions supported_ext = {
+      .float64 = device->instance->physicalDevice.info.gen >= 8,
+      .int64 = device->instance->physicalDevice.info.gen >= 8,
+      .tessellation = true,
+      .draw_parameters = true,
+      .image_write_without_format = true,
+   };
+
    nir_function *entry_point =
       spirv_to_nir(spirv, module->size / 4,
                    spec_entries, num_spec_entries,
-                   stage, entrypoint_name, nir_options);
+                   stage, entrypoint_name, &supported_ext, nir_options);
    nir_shader *nir = entry_point->shader;
    assert(nir->stage == stage);
    nir_validate_shader(nir);
 
    free(spec_entries);
 
-   if (stage == MESA_SHADER_FRAGMENT) {
-      nir_lower_wpos_center(nir);
-      nir_validate_shader(nir);
-   }
-
-   nir_lower_returns(nir);
-   nir_validate_shader(nir);
-
-   nir_inline_functions(nir);
-   nir_validate_shader(nir);
+   /* We have to lower away local constant initializers right before we
+    * inline functions.  That way they get properly initialized at the top
+    * of the function and not at the top of its caller.
+    */
+   NIR_PASS_V(nir, nir_lower_constant_initializers, nir_var_local);
+   NIR_PASS_V(nir, nir_lower_returns);
+   NIR_PASS_V(nir, nir_inline_functions);
 
    /* Pick off the single entrypoint that we want */
    foreach_list_typed_safe(nir_function, func, node, &nir->functions) {
@@ -147,36 +156,30 @@ anv_shader_compile_to_nir(struct anv_device *device,
    assert(exec_list_length(&nir->functions) == 1);
    entry_point->name = ralloc_strdup(entry_point, "main");
 
-   nir_remove_dead_variables(nir, nir_var_shader_in);
-   nir_remove_dead_variables(nir, nir_var_shader_out);
-   nir_remove_dead_variables(nir, nir_var_system_value);
-   nir_validate_shader(nir);
-
-   nir_propagate_invariant(nir);
-   nir_validate_shader(nir);
+   NIR_PASS_V(nir, nir_remove_dead_variables,
+              nir_var_shader_in | nir_var_shader_out | nir_var_system_value);
 
-   nir_lower_io_to_temporaries(entry_point->shader, entry_point->impl,
-                               true, false);
+   if (stage == MESA_SHADER_FRAGMENT)
+      NIR_PASS_V(nir, nir_lower_wpos_center, pipeline->sample_shading_enable);
 
-   nir_lower_system_values(nir);
-   nir_validate_shader(nir);
+   /* Now that we've deleted all but the main function, we can go ahead and
+    * lower the rest of the constant initializers.
+    */
+   NIR_PASS_V(nir, nir_lower_constant_initializers, ~0);
+   NIR_PASS_V(nir, nir_propagate_invariant);
+   NIR_PASS_V(nir, nir_lower_io_to_temporaries,
+              entry_point->impl, true, false);
+   NIR_PASS_V(nir, nir_lower_system_values);
 
    /* Vulkan uses the separate-shader linking model */
    nir->info->separate_shader = true;
 
    nir = brw_preprocess_nir(compiler, nir);
 
-   nir_shader_gather_info(nir, entry_point->impl);
-
-   nir_variable_mode indirect_mask = 0;
-   if (compiler->glsl_compiler_options[stage].EmitNoIndirectInput)
-      indirect_mask |= nir_var_shader_in;
-   if (compiler->glsl_compiler_options[stage].EmitNoIndirectOutput)
-      indirect_mask |= nir_var_shader_out;
-   if (compiler->glsl_compiler_options[stage].EmitNoIndirectTemp)
-      indirect_mask |= nir_var_local;
+   NIR_PASS_V(nir, nir_lower_clip_cull_distance_arrays);
 
-   nir_lower_indirect_derefs(nir, indirect_mask);
+   if (stage == MESA_SHADER_FRAGMENT)
+      NIR_PASS_V(nir, anv_nir_lower_input_attachments);
 
    return nir;
 }
@@ -189,6 +192,9 @@ void anv_DestroyPipeline(
    ANV_FROM_HANDLE(anv_device, device, _device);
    ANV_FROM_HANDLE(anv_pipeline, pipeline, _pipeline);
 
+   if (!pipeline)
+      return;
+
    anv_reloc_list_finish(&pipeline->batch_relocs,
                          pAllocator ? pAllocator : &device->alloc);
    if (pipeline->blend_state.map)
@@ -213,13 +219,31 @@ static const uint32_t vk_to_gen_primitive_type[] = {
    [VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY]     = _3DPRIM_LINESTRIP_ADJ,
    [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY]  = _3DPRIM_TRILIST_ADJ,
    [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY] = _3DPRIM_TRISTRIP_ADJ,
-/*   [VK_PRIMITIVE_TOPOLOGY_PATCH_LIST]                = _3DPRIM_PATCHLIST_1 */
 };
 
 static void
 populate_sampler_prog_key(const struct gen_device_info *devinfo,
                           struct brw_sampler_prog_key_data *key)
 {
+   /* Almost all multisampled textures are compressed.  The only time when we
+    * don't compress a multisampled texture is for 16x MSAA with a surface
+    * width greater than 8k which is a bit of an edge case.  Since the sampler
+    * just ignores the MCS parameter to ld2ms when MCS is disabled, it's safe
+    * to tell the compiler to always assume compression.
+    */
+   key->compressed_multisample_layout_mask = ~0;
+
+   /* SkyLake added support for 16x MSAA.  With this came a new message for
+    * reading from a 16x MSAA surface with compression.  The new message was
+    * needed because now the MCS data is 64 bits instead of 32 or lower as is
+    * the case for 8x, 4x, and 2x.  The key->msaa_16 bit-field controls which
+    * message we use.  Fortunately, the 16x message works for 8x, 4x, and 2x
+    * so we can just use it unconditionally.  This may not be quite as
+    * efficient but it saves us from recompiling.
+    */
+   if (devinfo->gen >= 9)
+      key->msaa_16 = ~0;
+
    /* XXX: Handle texture swizzle on HSW- */
    for (int i = 0; i < MAX_SAMPLERS; i++) {
       /* Assume color sampler, no swizzling. (Works for BDW+) */
@@ -250,17 +274,22 @@ populate_gs_prog_key(const struct gen_device_info *devinfo,
 }
 
 static void
-populate_wm_prog_key(const struct gen_device_info *devinfo,
+populate_wm_prog_key(const struct anv_pipeline *pipeline,
                      const VkGraphicsPipelineCreateInfo *info,
                      struct brw_wm_prog_key *key)
 {
+   const struct gen_device_info *devinfo = &pipeline->device->info;
    ANV_FROM_HANDLE(anv_render_pass, render_pass, info->renderPass);
 
    memset(key, 0, sizeof(*key));
 
    populate_sampler_prog_key(devinfo, &key->tex);
 
-   /* TODO: Fill out key->input_slots_valid */
+   /* TODO: we could set this to 0 based on the information in nir_shader, but
+    * this function is called before spirv_to_nir. */
+   const struct brw_vue_map *vue_map =
+      &anv_pipeline_get_last_vue_prog_data(pipeline)->vue_map;
+   key->input_slots_valid = vue_map->slots_valid;
 
    /* Vulkan doesn't specify a default */
    key->high_quality_derivatives = false;
@@ -275,14 +304,19 @@ populate_wm_prog_key(const struct gen_device_info *devinfo,
                           info->pMultisampleState &&
                           info->pMultisampleState->alphaToCoverageEnable;
 
-   if (info->pMultisampleState && info->pMultisampleState->rasterizationSamples > 1) {
+   if (info->pMultisampleState) {
       /* We should probably pull this out of the shader, but it's fairly
        * harmless to compute it and then let dead-code take care of it.
        */
-      key->persample_interp =
-         (info->pMultisampleState->minSampleShading *
-          info->pMultisampleState->rasterizationSamples) > 1;
-      key->multisample_fbo = true;
+      if (info->pMultisampleState->rasterizationSamples > 1) {
+         key->persample_interp =
+            (info->pMultisampleState->minSampleShading *
+             info->pMultisampleState->rasterizationSamples) > 1;
+         key->multisample_fbo = true;
+      }
+
+      key->frag_coord_adds_sample_pos =
+         info->pMultisampleState->sampleShadingEnable;
    }
 }
 
@@ -295,6 +329,34 @@ populate_cs_prog_key(const struct gen_device_info *devinfo,
    populate_sampler_prog_key(devinfo, &key->tex);
 }
 
+static void
+anv_pipeline_hash_shader(struct anv_pipeline *pipeline,
+                         struct anv_shader_module *module,
+                         const char *entrypoint,
+                         gl_shader_stage stage,
+                         const VkSpecializationInfo *spec_info,
+                         const void *key, size_t key_size,
+                         unsigned char *sha1_out)
+{
+   struct mesa_sha1 ctx;
+
+   _mesa_sha1_init(&ctx);
+   if (pipeline->layout) {
+      _mesa_sha1_update(&ctx, pipeline->layout->sha1,
+                        sizeof(pipeline->layout->sha1));
+   }
+   _mesa_sha1_update(&ctx, module->sha1, sizeof(module->sha1));
+   _mesa_sha1_update(&ctx, entrypoint, strlen(entrypoint));
+   _mesa_sha1_update(&ctx, &stage, sizeof(stage));
+   if (spec_info) {
+      _mesa_sha1_update(&ctx, spec_info->pMapEntries,
+                        spec_info->mapEntryCount * sizeof(*spec_info->pMapEntries));
+      _mesa_sha1_update(&ctx, spec_info->pData, spec_info->dataSize);
+   }
+   _mesa_sha1_update(&ctx, key, key_size);
+   _mesa_sha1_final(&ctx, sha1_out);
+}
+
 static nir_shader *
 anv_pipeline_compile(struct anv_pipeline *pipeline,
                      struct anv_shader_module *module,
@@ -304,13 +366,15 @@ anv_pipeline_compile(struct anv_pipeline *pipeline,
                      struct brw_stage_prog_data *prog_data,
                      struct anv_pipeline_bind_map *map)
 {
-   nir_shader *nir = anv_shader_compile_to_nir(pipeline->device,
+   nir_shader *nir = anv_shader_compile_to_nir(pipeline,
                                                module, entrypoint, stage,
                                                spec_info);
    if (nir == NULL)
       return NULL;
 
-   anv_nir_lower_push_constants(nir);
+   NIR_PASS_V(nir, anv_nir_lower_push_constants);
+
+   nir_shader_gather_info(nir, nir_shader_get_entrypoint(nir));
 
    /* Figure out the number of parameters */
    prog_data->nr_params = 0;
@@ -323,9 +387,6 @@ anv_pipeline_compile(struct anv_pipeline *pipeline,
       prog_data->nr_params += MAX_PUSH_CONSTANTS_SIZE / sizeof(float);
    }
 
-   if (pipeline->layout && pipeline->layout->stage[stage].has_dynamic_offsets)
-      prog_data->nr_params += MAX_DYNAMIC_BUFFERS * 2;
-
    if (nir->info->num_images > 0) {
       prog_data->nr_params += nir->info->num_images * BRW_IMAGE_PARAM_SIZE;
       pipeline->needs_data_cache = true;
@@ -357,9 +418,6 @@ anv_pipeline_compile(struct anv_pipeline *pipeline,
       }
    }
 
-   /* Set up dynamic offsets */
-   anv_nir_apply_dynamic_offsets(pipeline, nir, prog_data);
-
    /* Apply the actual pipeline layout to UBOs, SSBOs, and textures */
    if (pipeline->layout)
       anv_nir_apply_pipeline_layout(pipeline, nir, prog_data, map);
@@ -433,8 +491,9 @@ anv_pipeline_compile_vs(struct anv_pipeline *pipeline,
    populate_vs_prog_key(&pipeline->device->info, &key);
 
    if (cache) {
-      anv_hash_shader(sha1, &key, sizeof(key), module, entrypoint,
-                      pipeline->layout, spec_info);
+      anv_pipeline_hash_shader(pipeline, module, entrypoint,
+                               MESA_SHADER_VERTEX, spec_info,
+                               &key, sizeof(key), sha1);
       bin = anv_pipeline_cache_search(cache, sha1, 20);
    }
 
@@ -461,6 +520,7 @@ anv_pipeline_compile_vs(struct anv_pipeline *pipeline,
       ralloc_steal(mem_ctx, nir);
 
       prog_data.inputs_read = nir->info->inputs_read;
+      prog_data.double_inputs_read = nir->info->double_inputs_read;
 
       brw_compute_vue_map(&pipeline->device->info,
                           &prog_data.base.vue_map,
@@ -488,18 +548,191 @@ anv_pipeline_compile_vs(struct anv_pipeline *pipeline,
       ralloc_free(mem_ctx);
    }
 
-   const struct brw_vs_prog_data *vs_prog_data =
-      (const struct brw_vs_prog_data *)bin->prog_data;
+   anv_pipeline_add_compiled_stage(pipeline, MESA_SHADER_VERTEX, bin);
 
-   if (vs_prog_data->base.dispatch_mode == DISPATCH_MODE_SIMD8) {
-      pipeline->vs_simd8 = bin->kernel.offset;
-      pipeline->vs_vec4 = NO_KERNEL;
-   } else {
-      pipeline->vs_simd8 = NO_KERNEL;
-      pipeline->vs_vec4 = bin->kernel.offset;
+   return VK_SUCCESS;
+}
+
+static void
+merge_tess_info(struct shader_info *tes_info,
+                const struct shader_info *tcs_info)
+{
+   /* The Vulkan 1.0.38 spec, section 21.1 Tessellator says:
+    *
+    *    "PointMode. Controls generation of points rather than triangles
+    *     or lines. This functionality defaults to disabled, and is
+    *     enabled if either shader stage includes the execution mode.
+    *
+    * and about Triangles, Quads, IsoLines, VertexOrderCw, VertexOrderCcw,
+    * PointMode, SpacingEqual, SpacingFractionalEven, SpacingFractionalOdd,
+    * and OutputVertices, it says:
+    *
+    *    "One mode must be set in at least one of the tessellation
+    *     shader stages."
+    *
+    * So, the fields can be set in either the TCS or TES, but they must
+    * agree if set in both.  Our backend looks at TES, so bitwise-or in
+    * the values from the TCS.
+    */
+   assert(tcs_info->tess.tcs_vertices_out == 0 ||
+          tes_info->tess.tcs_vertices_out == 0 ||
+          tcs_info->tess.tcs_vertices_out == tes_info->tess.tcs_vertices_out);
+   tes_info->tess.tcs_vertices_out |= tcs_info->tess.tcs_vertices_out;
+
+   assert(tcs_info->tess.spacing == TESS_SPACING_UNSPECIFIED ||
+          tes_info->tess.spacing == TESS_SPACING_UNSPECIFIED ||
+          tcs_info->tess.spacing == tes_info->tess.spacing);
+   tes_info->tess.spacing |= tcs_info->tess.spacing;
+
+   tes_info->tess.ccw |= tcs_info->tess.ccw;
+   tes_info->tess.point_mode |= tcs_info->tess.point_mode;
+}
+
+static VkResult
+anv_pipeline_compile_tcs_tes(struct anv_pipeline *pipeline,
+                             struct anv_pipeline_cache *cache,
+                             const VkGraphicsPipelineCreateInfo *info,
+                             struct anv_shader_module *tcs_module,
+                             const char *tcs_entrypoint,
+                             const VkSpecializationInfo *tcs_spec_info,
+                             struct anv_shader_module *tes_module,
+                             const char *tes_entrypoint,
+                             const VkSpecializationInfo *tes_spec_info)
+{
+   const struct gen_device_info *devinfo = &pipeline->device->info;
+   const struct brw_compiler *compiler =
+      pipeline->device->instance->physicalDevice.compiler;
+   struct anv_pipeline_bind_map tcs_map;
+   struct anv_pipeline_bind_map tes_map;
+   struct brw_tcs_prog_key tcs_key = { 0, };
+   struct brw_tes_prog_key tes_key = { 0, };
+   struct anv_shader_bin *tcs_bin = NULL;
+   struct anv_shader_bin *tes_bin = NULL;
+   unsigned char tcs_sha1[40];
+   unsigned char tes_sha1[40];
+
+   populate_sampler_prog_key(&pipeline->device->info, &tcs_key.tex);
+   populate_sampler_prog_key(&pipeline->device->info, &tes_key.tex);
+   tcs_key.input_vertices = info->pTessellationState->patchControlPoints;
+
+   if (cache) {
+      anv_pipeline_hash_shader(pipeline, tcs_module, tcs_entrypoint,
+                               MESA_SHADER_TESS_CTRL, tcs_spec_info,
+                               &tcs_key, sizeof(tcs_key), tcs_sha1);
+      anv_pipeline_hash_shader(pipeline, tes_module, tes_entrypoint,
+                               MESA_SHADER_TESS_EVAL, tes_spec_info,
+                               &tes_key, sizeof(tes_key), tes_sha1);
+      memcpy(&tcs_sha1[20], tes_sha1, 20);
+      memcpy(&tes_sha1[20], tcs_sha1, 20);
+      tcs_bin = anv_pipeline_cache_search(cache, tcs_sha1, sizeof(tcs_sha1));
+      tes_bin = anv_pipeline_cache_search(cache, tes_sha1, sizeof(tes_sha1));
    }
 
-   anv_pipeline_add_compiled_stage(pipeline, MESA_SHADER_VERTEX, bin);
+   if (tcs_bin == NULL || tes_bin == NULL) {
+      struct brw_tcs_prog_data tcs_prog_data = { 0, };
+      struct brw_tes_prog_data tes_prog_data = { 0, };
+      struct anv_pipeline_binding tcs_surface_to_descriptor[256];
+      struct anv_pipeline_binding tcs_sampler_to_descriptor[256];
+      struct anv_pipeline_binding tes_surface_to_descriptor[256];
+      struct anv_pipeline_binding tes_sampler_to_descriptor[256];
+
+      tcs_map = (struct anv_pipeline_bind_map) {
+         .surface_to_descriptor = tcs_surface_to_descriptor,
+         .sampler_to_descriptor = tcs_sampler_to_descriptor
+      };
+      tes_map = (struct anv_pipeline_bind_map) {
+         .surface_to_descriptor = tes_surface_to_descriptor,
+         .sampler_to_descriptor = tes_sampler_to_descriptor
+      };
+
+      nir_shader *tcs_nir =
+         anv_pipeline_compile(pipeline, tcs_module, tcs_entrypoint,
+                              MESA_SHADER_TESS_CTRL, tcs_spec_info,
+                              &tcs_prog_data.base.base, &tcs_map);
+      nir_shader *tes_nir =
+         anv_pipeline_compile(pipeline, tes_module, tes_entrypoint,
+                              MESA_SHADER_TESS_EVAL, tes_spec_info,
+                              &tes_prog_data.base.base, &tes_map);
+      if (tcs_nir == NULL || tes_nir == NULL)
+         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+      nir_lower_tes_patch_vertices(tes_nir,
+                                   tcs_nir->info->tess.tcs_vertices_out);
+
+      /* Copy TCS info into the TES info */
+      merge_tess_info(tes_nir->info, tcs_nir->info);
+
+      anv_fill_binding_table(&tcs_prog_data.base.base, 0);
+      anv_fill_binding_table(&tes_prog_data.base.base, 0);
+
+      void *mem_ctx = ralloc_context(NULL);
+
+      ralloc_steal(mem_ctx, tcs_nir);
+      ralloc_steal(mem_ctx, tes_nir);
+
+      /* Whacking the key after cache lookup is a bit sketchy, but all of
+       * this comes from the SPIR-V, which is part of the hash used for the
+       * pipeline cache.  So it should be safe.
+       */
+      tcs_key.tes_primitive_mode = tes_nir->info->tess.primitive_mode;
+      tcs_key.outputs_written = tcs_nir->info->outputs_written;
+      tcs_key.patch_outputs_written = tcs_nir->info->patch_outputs_written;
+      tcs_key.quads_workaround =
+         devinfo->gen < 9 &&
+         tes_nir->info->tess.primitive_mode == 7 /* GL_QUADS */ &&
+         tes_nir->info->tess.spacing == TESS_SPACING_EQUAL;
+
+      tes_key.inputs_read = tcs_key.outputs_written;
+      tes_key.patch_inputs_read = tcs_key.patch_outputs_written;
+
+      unsigned code_size;
+      const int shader_time_index = -1;
+      const unsigned *shader_code;
+
+      shader_code =
+         brw_compile_tcs(compiler, NULL, mem_ctx, &tcs_key, &tcs_prog_data,
+                         tcs_nir, shader_time_index, &code_size, NULL);
+      if (shader_code == NULL) {
+         ralloc_free(mem_ctx);
+         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+      }
+
+      tcs_bin = anv_pipeline_upload_kernel(pipeline, cache,
+                                           tcs_sha1, sizeof(tcs_sha1),
+                                           shader_code, code_size,
+                                           &tcs_prog_data.base.base,
+                                           sizeof(tcs_prog_data),
+                                           &tcs_map);
+      if (!tcs_bin) {
+         ralloc_free(mem_ctx);
+         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+      }
+
+      shader_code =
+         brw_compile_tes(compiler, NULL, mem_ctx, &tes_key,
+                         &tcs_prog_data.base.vue_map, &tes_prog_data, tes_nir,
+                         NULL, shader_time_index, &code_size, NULL);
+      if (shader_code == NULL) {
+         ralloc_free(mem_ctx);
+         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+      }
+
+      tes_bin = anv_pipeline_upload_kernel(pipeline, cache,
+                                           tes_sha1, sizeof(tes_sha1),
+                                           shader_code, code_size,
+                                           &tes_prog_data.base.base,
+                                           sizeof(tes_prog_data),
+                                           &tes_map);
+      if (!tes_bin) {
+         ralloc_free(mem_ctx);
+         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+      }
+
+      ralloc_free(mem_ctx);
+   }
+
+   anv_pipeline_add_compiled_stage(pipeline, MESA_SHADER_TESS_CTRL, tcs_bin);
+   anv_pipeline_add_compiled_stage(pipeline, MESA_SHADER_TESS_EVAL, tes_bin);
 
    return VK_SUCCESS;
 }
@@ -522,8 +755,9 @@ anv_pipeline_compile_gs(struct anv_pipeline *pipeline,
    populate_gs_prog_key(&pipeline->device->info, &key);
 
    if (cache) {
-      anv_hash_shader(sha1, &key, sizeof(key), module, entrypoint,
-                      pipeline->layout, spec_info);
+      anv_pipeline_hash_shader(pipeline, module, entrypoint,
+                               MESA_SHADER_GEOMETRY, spec_info,
+                               &key, sizeof(key), sha1);
       bin = anv_pipeline_cache_search(cache, sha1, 20);
    }
 
@@ -576,8 +810,6 @@ anv_pipeline_compile_gs(struct anv_pipeline *pipeline,
       ralloc_free(mem_ctx);
    }
 
-   pipeline->gs_kernel = bin->kernel.offset;
-
    anv_pipeline_add_compiled_stage(pipeline, MESA_SHADER_GEOMETRY, bin);
 
    return VK_SUCCESS;
@@ -598,11 +830,12 @@ anv_pipeline_compile_fs(struct anv_pipeline *pipeline,
    struct anv_shader_bin *bin = NULL;
    unsigned char sha1[20];
 
-   populate_wm_prog_key(&pipeline->device->info, info, &key);
+   populate_wm_prog_key(pipeline, info, &key);
 
    if (cache) {
-      anv_hash_shader(sha1, &key, sizeof(key), module, entrypoint,
-                      pipeline->layout, spec_info);
+      anv_pipeline_hash_shader(pipeline, module, entrypoint,
+                               MESA_SHADER_FRAGMENT, spec_info,
+                               &key, sizeof(key), sha1);
       bin = anv_pipeline_cache_search(cache, sha1, 20);
    }
 
@@ -700,8 +933,6 @@ anv_pipeline_compile_fs(struct anv_pipeline *pipeline,
       ralloc_free(mem_ctx);
    }
 
-   pipeline->ps_ksp0 = bin->kernel.offset;
-
    anv_pipeline_add_compiled_stage(pipeline, MESA_SHADER_FRAGMENT, bin);
 
    return VK_SUCCESS;
@@ -725,8 +956,9 @@ anv_pipeline_compile_cs(struct anv_pipeline *pipeline,
    populate_cs_prog_key(&pipeline->device->info, &key);
 
    if (cache) {
-      anv_hash_shader(sha1, &key, sizeof(key), module, entrypoint,
-                      pipeline->layout, spec_info);
+      anv_pipeline_hash_shader(pipeline, module, entrypoint,
+                               MESA_SHADER_COMPUTE, spec_info,
+                               &key, sizeof(key), sha1);
       bin = anv_pipeline_cache_search(cache, sha1, 20);
    }
 
@@ -773,8 +1005,6 @@ anv_pipeline_compile_cs(struct anv_pipeline *pipeline,
       ralloc_free(mem_ctx);
    }
 
-   pipeline->cs_simd = bin->kernel.offset;
-
    anv_pipeline_add_compiled_stage(pipeline, MESA_SHADER_COMPUTE, bin);
 
    return VK_SUCCESS;
@@ -858,7 +1088,7 @@ copy_non_dynamic_state(struct anv_pipeline *pipeline,
     */
    bool uses_color_att = false;
    for (unsigned i = 0; i < subpass->color_count; ++i) {
-      if (subpass->color_attachments[i] != VK_ATTACHMENT_UNUSED) {
+      if (subpass->color_attachments[i].attachment != VK_ATTACHMENT_UNUSED) {
          uses_color_att = true;
          break;
       }
@@ -886,7 +1116,7 @@ copy_non_dynamic_state(struct anv_pipeline *pipeline,
     *    against does not use a depth/stencil attachment.
     */
    if (!pCreateInfo->pRasterizationState->rasterizerDiscardEnable &&
-       subpass->depth_stencil_attachment != VK_ATTACHMENT_UNUSED) {
+       subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED) {
       assert(pCreateInfo->pDepthStencilState);
 
       if (states & (1 << VK_DYNAMIC_STATE_DEPTH_BOUNDS)) {
@@ -924,6 +1154,7 @@ copy_non_dynamic_state(struct anv_pipeline *pipeline,
 static void
 anv_pipeline_validate_create_info(const VkGraphicsPipelineCreateInfo *info)
 {
+#ifdef DEBUG
    struct anv_render_pass *renderpass = NULL;
    struct anv_subpass *subpass = NULL;
 
@@ -946,7 +1177,7 @@ anv_pipeline_validate_create_info(const VkGraphicsPipelineCreateInfo *info)
       assert(info->pViewportState);
       assert(info->pMultisampleState);
 
-      if (subpass && subpass->depth_stencil_attachment != VK_ATTACHMENT_UNUSED)
+      if (subpass && subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED)
          assert(info->pDepthStencilState);
 
       if (subpass && subpass->color_count > 0)
@@ -963,6 +1194,7 @@ anv_pipeline_validate_create_info(const VkGraphicsPipelineCreateInfo *info)
          break;
       }
    }
+#endif
 }
 
 /**
@@ -994,9 +1226,7 @@ anv_pipeline_init(struct anv_pipeline *pipeline,
 {
    VkResult result;
 
-   anv_validate {
-      anv_pipeline_validate_create_info(pCreateInfo);
-   }
+   anv_pipeline_validate_create_info(pCreateInfo);
 
    if (alloc == NULL)
       alloc = &device->alloc;
@@ -1012,11 +1242,15 @@ anv_pipeline_init(struct anv_pipeline *pipeline,
    pipeline->batch.next = pipeline->batch.start = pipeline->batch_data;
    pipeline->batch.end = pipeline->batch.start + sizeof(pipeline->batch_data);
    pipeline->batch.relocs = &pipeline->batch_relocs;
+   pipeline->batch.status = VK_SUCCESS;
 
    copy_non_dynamic_state(pipeline, pCreateInfo);
    pipeline->depth_clamp_enable = pCreateInfo->pRasterizationState &&
                                   pCreateInfo->pRasterizationState->depthClampEnable;
 
+   pipeline->sample_shading_enable = pCreateInfo->pMultisampleState &&
+                                     pCreateInfo->pMultisampleState->sampleShadingEnable;
+
    pipeline->needs_data_cache = false;
 
    /* When we free the pipeline, we detect stages based on the NULL status
@@ -1024,11 +1258,6 @@ anv_pipeline_init(struct anv_pipeline *pipeline,
     */
    memset(pipeline->shaders, 0, sizeof(pipeline->shaders));
 
-   pipeline->vs_simd8 = NO_KERNEL;
-   pipeline->vs_vec4 = NO_KERNEL;
-   pipeline->gs_kernel = NO_KERNEL;
-   pipeline->ps_ksp0 = NO_KERNEL;
-
    pipeline->active_stages = 0;
 
    const VkPipelineShaderStageCreateInfo *pStages[MESA_SHADER_STAGES] = { 0, };
@@ -1048,8 +1277,15 @@ anv_pipeline_init(struct anv_pipeline *pipeline,
          goto compile_fail;
    }
 
-   if (modules[MESA_SHADER_TESS_CTRL] || modules[MESA_SHADER_TESS_EVAL])
-      anv_finishme("no tessellation support");
+   if (modules[MESA_SHADER_TESS_EVAL]) {
+      anv_pipeline_compile_tcs_tes(pipeline, cache, pCreateInfo,
+                                   modules[MESA_SHADER_TESS_CTRL],
+                                   pStages[MESA_SHADER_TESS_CTRL]->pName,
+                                   pStages[MESA_SHADER_TESS_CTRL]->pSpecializationInfo,
+                                   modules[MESA_SHADER_TESS_EVAL],
+                                   pStages[MESA_SHADER_TESS_EVAL]->pName,
+                                   pStages[MESA_SHADER_TESS_EVAL]->pSpecializationInfo);
+   }
 
    if (modules[MESA_SHADER_GEOMETRY]) {
       result = anv_pipeline_compile_gs(pipeline, cache, pCreateInfo,
@@ -1109,8 +1345,14 @@ anv_pipeline_init(struct anv_pipeline *pipeline,
 
    const VkPipelineInputAssemblyStateCreateInfo *ia_info =
       pCreateInfo->pInputAssemblyState;
+   const VkPipelineTessellationStateCreateInfo *tess_info =
+      pCreateInfo->pTessellationState;
    pipeline->primitive_restart = ia_info->primitiveRestartEnable;
-   pipeline->topology = vk_to_gen_primitive_type[ia_info->topology];
+
+   if (anv_pipeline_has_stage(pipeline, MESA_SHADER_TESS_EVAL))
+      pipeline->topology = _3DPRIM_PATCHLIST(tess_info->patchControlPoints);
+   else
+      pipeline->topology = vk_to_gen_primitive_type[ia_info->topology];
 
    return VK_SUCCESS;