OSDN Git Service

turnip: preliminary support for shader modules
authorChia-I Wu <olvaffe@gmail.com>
Wed, 20 Feb 2019 17:53:47 +0000 (09:53 -0800)
committerChia-I Wu <olvaffe@gmail.com>
Mon, 11 Mar 2019 17:02:13 +0000 (10:02 -0700)
Save SPIR-V in tu_shader_module.  Tranlation to NIR happens in
tu_shader_create, and compilation to binary code happens in
tu_shader_compile.  Both will be called during pipeline creation.

src/freedreno/vulkan/.clang-format
src/freedreno/vulkan/meson.build
src/freedreno/vulkan/tu_device.c
src/freedreno/vulkan/tu_private.h
src/freedreno/vulkan/tu_shader.c [new file with mode: 0644]

index 7ff44c5..c6bf5df 100644 (file)
@@ -16,7 +16,7 @@ IncludeBlocks: Regroup
 IncludeCategories:
   - Regex:           '^"tu_private.h"$'
     Priority:        0
-  - Regex:           '^"(drm/|tu_)'
+  - Regex:           '^"(drm/|ir3/|tu_)'
     Priority:        4
   - Regex:           '^"(c11/|compiler/|main/|nir/|spirv/|util/|vk_)'
     Priority:        3
index 3b2624e..6c0f7d9 100644 (file)
@@ -68,6 +68,7 @@ libtu_files = files(
   'tu_pipeline_cache.c',
   'tu_private.h',
   'tu_query.c',
+  'tu_shader.c',
   'tu_util.c',
   'tu_util.h',
   'vk_format.h',
@@ -88,6 +89,8 @@ libvulkan_freedreno = shared_library(
   link_with : [
     libvulkan_util,
     libmesa_util,
+    libfreedreno_drm, # required by ir3_shader_get_variant, which we don't use
+    libfreedreno_ir3,
   ],
   dependencies : [
     dep_dl,
index 76d0f21..e58865e 100644 (file)
@@ -344,7 +344,10 @@ static const VkAllocationCallbacks default_alloc = {
 };
 
 static const struct debug_control tu_debug_options[] = {
-   { "startup", TU_DEBUG_STARTUP }, { NULL, 0 }
+   { "startup", TU_DEBUG_STARTUP },
+   { "nir", TU_DEBUG_NIR },
+   { "ir3", TU_DEBUG_IR3 },
+   { NULL, 0 }
 };
 
 const char *
@@ -1078,6 +1081,10 @@ tu_CreateDevice(VkPhysicalDevice physicalDevice,
       }
    }
 
+   device->compiler = ir3_compiler_create(NULL, physical_device->gpu_id);
+   if (!device->compiler)
+      goto fail;
+
    VkPipelineCacheCreateInfo ci;
    ci.sType = VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO;
    ci.pNext = NULL;
@@ -1103,6 +1110,9 @@ fail:
          vk_free(&device->alloc, device->queues[i]);
    }
 
+   if (device->compiler)
+      ralloc_free(device->compiler);
+
    vk_free(&device->alloc, device);
    return result;
 }
@@ -1122,6 +1132,9 @@ tu_DestroyDevice(VkDevice _device, const VkAllocationCallbacks *pAllocator)
          vk_free(&device->alloc, device->queues[i]);
    }
 
+   /* the compiler does not use pAllocator */
+   ralloc_free(device->compiler);
+
    VkPipelineCache pc = tu_pipeline_cache_to_handle(device->mem_cache);
    tu_DestroyPipelineCache(tu_device_to_handle(device), pc, NULL);
 
index fa38a0f..57678e3 100644 (file)
@@ -52,6 +52,8 @@
 #include "vk_debug_report.h"
 
 #include "drm/msm_drm.h"
+#include "ir3/ir3_compiler.h"
+#include "ir3/ir3_shader.h"
 
 #include "adreno_common.xml.h"
 #include "adreno_pm4.xml.h"
@@ -315,6 +317,8 @@ struct tu_physical_device
 enum tu_debug_flags
 {
    TU_DEBUG_STARTUP = 1 << 0,
+   TU_DEBUG_NIR = 1 << 1,
+   TU_DEBUG_IR3 = 1 << 2,
 };
 
 struct tu_instance
@@ -447,6 +451,8 @@ struct tu_device
 
    struct tu_physical_device *physical_device;
 
+   struct ir3_compiler *compiler;
+
    /* Backup in-memory cache to be used if the app doesn't provide one */
    struct tu_pipeline_cache *mem_cache;
 
@@ -969,12 +975,58 @@ mesa_to_vk_shader_stage(gl_shader_stage mesa_stage)
 
 struct tu_shader_module
 {
-   struct nir_shader *nir;
    unsigned char sha1[20];
-   uint32_t size;
-   char data[0];
+
+   uint32_t code_size;
+   const uint32_t *code[0];
 };
 
+struct tu_shader_compile_options
+{
+   struct ir3_shader_key key;
+
+   bool optimize;
+   bool include_binning_pass;
+};
+
+struct tu_shader
+{
+   struct ir3_shader ir3_shader;
+
+   /* This may be true for vertex shaders.  When true, variants[1] is the
+    * binning variant and binning_binary is non-NULL.
+    */
+   bool has_binning_pass;
+
+   void *binary;
+   void *binning_binary;
+
+   struct ir3_shader_variant variants[0];
+};
+
+struct tu_shader *
+tu_shader_create(struct tu_device *dev,
+                 gl_shader_stage stage,
+                 const VkPipelineShaderStageCreateInfo *stage_info,
+                 const VkAllocationCallbacks *alloc);
+
+void
+tu_shader_destroy(struct tu_device *dev,
+                  struct tu_shader *shader,
+                  const VkAllocationCallbacks *alloc);
+
+void
+tu_shader_compile_options_init(
+   struct tu_shader_compile_options *options,
+   const VkGraphicsPipelineCreateInfo *pipeline_info);
+
+VkResult
+tu_shader_compile(struct tu_device *dev,
+                  struct tu_shader *shader,
+                  const struct tu_shader *next_stage,
+                  const struct tu_shader_compile_options *options,
+                  const VkAllocationCallbacks *alloc);
+
 struct tu_pipeline
 {
    struct tu_cs cs;
diff --git a/src/freedreno/vulkan/tu_shader.c b/src/freedreno/vulkan/tu_shader.c
new file mode 100644 (file)
index 0000000..2a70136
--- /dev/null
@@ -0,0 +1,338 @@
+/*
+ * Copyright © 2019 Google LLC
+ *
+ * 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 "tu_private.h"
+
+#include "spirv/nir_spirv.h"
+#include "util/mesa-sha1.h"
+
+#include "ir3/ir3_nir.h"
+
+static nir_function *
+tu_spirv_to_nir(struct ir3_compiler *compiler,
+                const uint32_t *words,
+                size_t word_count,
+                gl_shader_stage stage,
+                const char *entry_point_name,
+                const VkSpecializationInfo *spec_info)
+{
+   /* TODO these are made-up */
+   const struct spirv_to_nir_options spirv_options = {
+      .lower_workgroup_access_to_offsets = true,
+      .lower_ubo_ssbo_access_to_offsets = true,
+      .caps = { false },
+   };
+   const nir_shader_compiler_options *nir_options =
+      ir3_get_compiler_options(compiler);
+
+   /* convert VkSpecializationInfo */
+   struct nir_spirv_specialization *spec = NULL;
+   uint32_t num_spec = 0;
+   if (spec_info && spec_info->mapEntryCount) {
+      spec = malloc(sizeof(*spec) * spec_info->mapEntryCount);
+      if (!spec)
+         return NULL;
+
+      for (uint32_t i = 0; i < spec_info->mapEntryCount; i++) {
+         const VkSpecializationMapEntry *entry = &spec_info->pMapEntries[i];
+         const void *data = spec_info->pData + entry->offset;
+         assert(data + entry->size <= spec_info->pData + spec_info->dataSize);
+         spec[i].id = entry->constantID;
+         if (entry->size == 8)
+            spec[i].data64 = *(const uint64_t *) data;
+         else
+            spec[i].data32 = *(const uint32_t *) data;
+         spec[i].defined_on_module = false;
+      }
+
+      num_spec = spec_info->mapEntryCount;
+   }
+
+   nir_function *entry_point =
+      spirv_to_nir(words, word_count, spec, num_spec, stage, entry_point_name,
+                   &spirv_options, nir_options);
+
+   free(spec);
+
+   assert(entry_point->shader->info.stage == stage);
+   nir_validate_shader(entry_point->shader, "after spirv_to_nir");
+
+   return entry_point;
+}
+
+static void
+tu_sort_variables_by_location(struct exec_list *variables)
+{
+   struct exec_list sorted;
+   exec_list_make_empty(&sorted);
+
+   nir_foreach_variable_safe(var, variables)
+   {
+      exec_node_remove(&var->node);
+
+      /* insert the variable into the sorted list */
+      nir_variable *next = NULL;
+      nir_foreach_variable(tmp, &sorted)
+      {
+         if (var->data.location < tmp->data.location) {
+            next = tmp;
+            break;
+         }
+      }
+      if (next)
+         exec_node_insert_node_before(&next->node, &var->node);
+      else
+         exec_list_push_tail(&sorted, &var->node);
+   }
+
+   exec_list_move_nodes_to(&sorted, variables);
+}
+
+struct tu_shader *
+tu_shader_create(struct tu_device *dev,
+                 gl_shader_stage stage,
+                 const VkPipelineShaderStageCreateInfo *stage_info,
+                 const VkAllocationCallbacks *alloc)
+{
+   const struct tu_shader_module *module =
+      tu_shader_module_from_handle(stage_info->module);
+   struct tu_shader *shader;
+
+   const uint32_t max_variant_count = (stage == MESA_SHADER_VERTEX) ? 2 : 1;
+   shader = vk_zalloc2(
+      &dev->alloc, alloc,
+      sizeof(*shader) + sizeof(struct ir3_shader_variant) * max_variant_count,
+      8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND);
+   if (!shader)
+      return NULL;
+
+   /* translate SPIR-V to NIR */
+   assert(module->code_size % 4 == 0);
+   nir_function *entry_point = tu_spirv_to_nir(
+      dev->compiler, (const uint32_t *) module->code, module->code_size / 4,
+      stage, stage_info->pName, stage_info->pSpecializationInfo);
+   if (!entry_point) {
+      vk_free2(&dev->alloc, alloc, shader);
+      return NULL;
+   }
+
+   nir_shader *nir = entry_point->shader;
+
+   if (unlikely(dev->physical_device->instance->debug_flags & TU_DEBUG_NIR)) {
+      fprintf(stderr, "translated nir:\n");
+      nir_print_shader(nir, stderr);
+   }
+
+   /* TODO what needs to happen? */
+
+   switch (stage) {
+   case MESA_SHADER_VERTEX:
+      tu_sort_variables_by_location(&nir->outputs);
+      break;
+   case MESA_SHADER_TESS_CTRL:
+   case MESA_SHADER_TESS_EVAL:
+   case MESA_SHADER_GEOMETRY:
+      tu_sort_variables_by_location(&nir->inputs);
+      tu_sort_variables_by_location(&nir->outputs);
+      break;
+   case MESA_SHADER_FRAGMENT:
+      tu_sort_variables_by_location(&nir->inputs);
+      break;
+   case MESA_SHADER_COMPUTE:
+      break;
+   default:
+      unreachable("invalid gl_shader_stage");
+      break;
+   }
+
+   nir_assign_var_locations(&nir->inputs, &nir->num_inputs,
+                            ir3_glsl_type_size);
+   nir_assign_var_locations(&nir->outputs, &nir->num_outputs,
+                            ir3_glsl_type_size);
+   nir_assign_var_locations(&nir->uniforms, &nir->num_uniforms,
+                            ir3_glsl_type_size);
+
+   NIR_PASS_V(nir, nir_lower_system_values);
+   NIR_PASS_V(nir, nir_lower_io, nir_var_all, ir3_glsl_type_size, 0);
+
+   nir_shader_gather_info(nir, entry_point->impl);
+
+   shader->ir3_shader.compiler = dev->compiler;
+   shader->ir3_shader.type = stage;
+   shader->ir3_shader.nir = nir;
+
+   return shader;
+}
+
+void
+tu_shader_destroy(struct tu_device *dev,
+                  struct tu_shader *shader,
+                  const VkAllocationCallbacks *alloc)
+{
+   if (shader->ir3_shader.nir)
+      ralloc_free(shader->ir3_shader.nir);
+
+   for (uint32_t i = 0; i < 1 + shader->has_binning_pass; i++) {
+      if (shader->variants[i].ir)
+         ir3_destroy(shader->variants[i].ir);
+      if (shader->variants[i].immediates)
+         free(shader->variants[i].immediates);
+   }
+
+   if (shader->binary)
+      free(shader->binary);
+   if (shader->binning_binary)
+      free(shader->binning_binary);
+
+   vk_free2(&dev->alloc, alloc, shader);
+}
+
+void
+tu_shader_compile_options_init(
+   struct tu_shader_compile_options *options,
+   const VkGraphicsPipelineCreateInfo *pipeline_info)
+{
+   *options = (struct tu_shader_compile_options) {
+      /* TODO ir3_key */
+
+      .optimize = !(pipeline_info->flags &
+                    VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT),
+      .include_binning_pass = true,
+   };
+}
+
+static uint32_t *
+tu_compile_shader_variant(struct ir3_shader *shader,
+                          const struct ir3_shader_key *key,
+                          bool binning_pass,
+                          struct ir3_shader_variant *variant)
+{
+   variant->shader = shader;
+   variant->type = shader->type;
+   variant->key = *key;
+   variant->binning_pass = binning_pass;
+
+   int ret = ir3_compile_shader_nir(shader->compiler, variant);
+   if (ret)
+      return NULL;
+
+   /* when assemble fails, we rely on tu_shader_destroy to clean up the
+    * variant
+    */
+   return ir3_shader_assemble(variant, shader->compiler->gpu_id);
+}
+
+VkResult
+tu_shader_compile(struct tu_device *dev,
+                  struct tu_shader *shader,
+                  const struct tu_shader *next_stage,
+                  const struct tu_shader_compile_options *options,
+                  const VkAllocationCallbacks *alloc)
+{
+   if (options->optimize) {
+      /* ignore the key for the first pass of optimization */
+      ir3_optimize_nir(&shader->ir3_shader, shader->ir3_shader.nir, NULL);
+
+      if (unlikely(dev->physical_device->instance->debug_flags &
+                   TU_DEBUG_NIR)) {
+         fprintf(stderr, "optimized nir:\n");
+         nir_print_shader(shader->ir3_shader.nir, stderr);
+      }
+   }
+
+   shader->binary = tu_compile_shader_variant(
+      &shader->ir3_shader, &options->key, false, &shader->variants[0]);
+   if (!shader->binary)
+      return VK_ERROR_OUT_OF_HOST_MEMORY;
+
+   /* compile another variant for the binning pass */
+   if (options->include_binning_pass &&
+       shader->ir3_shader.type == MESA_SHADER_VERTEX) {
+      shader->binning_binary = tu_compile_shader_variant(
+         &shader->ir3_shader, &options->key, true, &shader->variants[1]);
+      if (!shader->binning_binary)
+         return VK_ERROR_OUT_OF_HOST_MEMORY;
+
+      shader->has_binning_pass = true;
+   }
+
+   if (unlikely(dev->physical_device->instance->debug_flags & TU_DEBUG_IR3)) {
+      fprintf(stderr, "disassembled ir3:\n");
+      fprintf(stderr, "shader: %s\n",
+              gl_shader_stage_name(shader->ir3_shader.type));
+      ir3_shader_disasm(&shader->variants[0], shader->binary, stderr);
+
+      if (shader->has_binning_pass) {
+         fprintf(stderr, "disassembled ir3:\n");
+         fprintf(stderr, "shader: %s (binning)\n",
+                 gl_shader_stage_name(shader->ir3_shader.type));
+         ir3_shader_disasm(&shader->variants[1], shader->binning_binary,
+                           stderr);
+      }
+   }
+
+   return VK_SUCCESS;
+}
+
+VkResult
+tu_CreateShaderModule(VkDevice _device,
+                      const VkShaderModuleCreateInfo *pCreateInfo,
+                      const VkAllocationCallbacks *pAllocator,
+                      VkShaderModule *pShaderModule)
+{
+   TU_FROM_HANDLE(tu_device, device, _device);
+   struct tu_shader_module *module;
+
+   assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO);
+   assert(pCreateInfo->flags == 0);
+   assert(pCreateInfo->codeSize % 4 == 0);
+
+   module = vk_alloc2(&device->alloc, pAllocator,
+                      sizeof(*module) + pCreateInfo->codeSize, 8,
+                      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+   if (module == NULL)
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
+
+   module->code_size = pCreateInfo->codeSize;
+   memcpy(module->code, pCreateInfo->pCode, pCreateInfo->codeSize);
+
+   _mesa_sha1_compute(module->code, module->code_size, module->sha1);
+
+   *pShaderModule = tu_shader_module_to_handle(module);
+
+   return VK_SUCCESS;
+}
+
+void
+tu_DestroyShaderModule(VkDevice _device,
+                       VkShaderModule _module,
+                       const VkAllocationCallbacks *pAllocator)
+{
+   TU_FROM_HANDLE(tu_device, device, _device);
+   TU_FROM_HANDLE(tu_shader_module, module, _module);
+
+   if (!module)
+      return;
+
+   vk_free2(&device->alloc, pAllocator, module);
+}