Tally/Readback cleanup;

src/api/l_graphics.c

@@ -680,6 +680,7 @@ static int l_lovrGraphicsGetFeatures(lua_State* L) {
   lua_pushboolean(L, features.wireframe), lua_setfield(L, -2, "wireframe");
   lua_pushboolean(L, features.depthClamp), lua_setfield(L, -2, "depthClamp");
   lua_pushboolean(L, features.indirectDrawFirstInstance), lua_setfield(L, -2, "indirectDrawFirstInstance");
+  lua_pushboolean(L, features.stageTally), lua_setfield(L, -2, "stageTally");
   lua_pushboolean(L, features.float64), lua_setfield(L, -2, "float64");
   lua_pushboolean(L, features.int64), lua_setfield(L, -2, "int64");
   lua_pushboolean(L, features.int16), lua_setfield(L, -2, "int16");

src/api/l_graphics_readback.c

@@ -1,5 +1,6 @@
 #include "api.h"
 #include "graphics/graphics.h"
+#include "data/blob.h"
 #include "data/image.h"
 #include "util.h"
 #include <lua.h>
@@ -35,7 +36,7 @@ static int l_lovrReadbackGetData(lua_State* L) {
       int count = (int) info->tally.count;
 
       if (lovrTallyGetInfo(info->tally.object)->type == TALLY_STAGE) {
-        count *= 6; // The number of pipeline statistics that are tracked
+        count *= 4; // The number of pipeline statistics that are tracked
       }
 
       lua_createtable(L, count, 0);
@@ -49,9 +50,9 @@ static int l_lovrReadbackGetData(lua_State* L) {
 }
 
 static int l_lovrReadbackGetBlob(lua_State* L) {
-  //Readback* readback = luax_checktype(L, 1, Readback);
-  //void* data = lovrReadbackGetData(readback);
-  lua_pushnil(L);
+  Readback* readback = luax_checktype(L, 1, Readback);
+  Blob* blob = lovrReadbackGetBlob(readback);
+  luax_pushtype(L, Blob, blob);
   return 1;
 }
 

src/core/gpu.h

@@ -454,7 +454,7 @@ void gpu_pipeline_destroy(gpu_pipeline* pipeline);
 typedef enum {
   GPU_TALLY_TIMER,
   GPU_TALLY_PIXEL,
-  GPU_TALLY_PIPELINE
+  GPU_TALLY_STAGE
 } gpu_tally_type;
 
 typedef struct {
@@ -604,6 +604,7 @@ typedef struct {
   bool wireframe;
   bool depthClamp;
   bool indirectDrawFirstInstance;
+  bool stageTally;
   bool float64;
   bool int64;
   bool int16;

src/core/gpu_vk.c

@@ -1367,14 +1367,14 @@ bool gpu_tally_init(gpu_tally* tally, gpu_tally_info* info) {
   VkQueryType queryTypes[] = {
     [GPU_TALLY_TIMER] = VK_QUERY_TYPE_TIMESTAMP,
     [GPU_TALLY_PIXEL] = VK_QUERY_TYPE_OCCLUSION,
-    [GPU_TALLY_PIPELINE] = VK_QUERY_TYPE_PIPELINE_STATISTICS
+    [GPU_TALLY_STAGE] = VK_QUERY_TYPE_PIPELINE_STATISTICS
   };
 
   VkQueryPoolCreateInfo createInfo = {
     .sType = VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO,
     .queryType = queryTypes[info->type],
     .queryCount = info->count,
-    .pipelineStatistics = info->type == GPU_TALLY_PIPELINE ? (
+    .pipelineStatistics = info->type == GPU_TALLY_STAGE ? (
       VK_QUERY_PIPELINE_STATISTIC_INPUT_ASSEMBLY_VERTICES_BIT |
       VK_QUERY_PIPELINE_STATISTIC_VERTEX_SHADER_INVOCATIONS_BIT |
       VK_QUERY_PIPELINE_STATISTIC_CLIPPING_PRIMITIVES_BIT |
@@ -1865,7 +1865,6 @@ bool gpu_init(gpu_config* config) {
       enable->shaderClipDistance = supports->shaderClipDistance;
       enable->shaderCullDistance = supports->shaderCullDistance;
       enable->largePoints = supports->largePoints;
-      enable->pipelineStatisticsQuery = supports->pipelineStatisticsQuery;
 
       // Optional features (currently always enabled when supported)
       config->features->textureBC = enable->textureCompressionBC = supports->textureCompressionBC;
@@ -1873,6 +1872,7 @@ bool gpu_init(gpu_config* config) {
       config->features->wireframe = enable->fillModeNonSolid = supports->fillModeNonSolid;
       config->features->depthClamp = enable->depthClamp = supports->depthClamp;
       config->features->indirectDrawFirstInstance = enable->drawIndirectFirstInstance = supports->drawIndirectFirstInstance;
+      config->features->stageTally = enable->pipelineStatisticsQuery = supports->pipelineStatisticsQuery;
       config->features->float64 = enable->shaderFloat64 = supports->shaderFloat64;
       config->features->int64 = enable->shaderInt64 = supports->shaderInt64;
       config->features->int16 = enable->shaderInt16 = supports->shaderInt16;

src/modules/graphics/graphics.c

@@ -222,6 +222,7 @@ struct Readback {
   gpu_buffer* buffer;
   void* pointer;
   Image* image;
+  Blob* blob;
   void* data;
 };
 
@@ -2734,6 +2735,7 @@ Readback* lovrReadbackCreate(const ReadbackInfo* info) {
       readback->size = info->buffer.extent;
       readback->data = malloc(readback->size);
       lovrAssert(readback->data, "Out of memory");
+      readback->blob = lovrBlobCreate(readback->data, readback->size, "Readback");
       break;
     case READBACK_TEXTURE:
       lovrRetain(info->texture.object);
@@ -2743,10 +2745,11 @@ Readback* lovrReadbackCreate(const ReadbackInfo* info) {
       break;
     case READBACK_TALLY:
       lovrRetain(info->tally.object);
-      uint32_t stride = info->tally.object->info.type == TALLY_STAGE ? 24 : 4;
+      uint32_t stride = info->tally.object->info.type == TALLY_STAGE ? 16 : 4;
       readback->size = info->tally.count * stride;
       readback->data = malloc(readback->size);
       lovrAssert(readback->data, "Out of memory");
+      readback->blob = lovrBlobCreate(readback->data, readback->size, "Readback");
       break;
   }
 
@@ -2773,7 +2776,7 @@ void lovrReadbackDestroy(void* ref) {
     case READBACK_TALLY: lovrRelease(readback->info.tally.object, lovrTallyDestroy); break;
   }
   lovrRelease(readback->image, lovrImageDestroy);
-  free(readback->data);
+  lovrRelease(readback->blob, lovrBlobDestroy);
   free(readback);
 }
 
@@ -2803,6 +2806,10 @@ void* lovrReadbackGetData(Readback* readback) {
   return lovrReadbackIsComplete(readback) ? readback->data : NULL;
 }
 
+Blob* lovrReadbackGetBlob(Readback* readback) {
+  return lovrReadbackIsComplete(readback) ? readback->blob : NULL;
+}
+
 Image* lovrReadbackGetImage(Readback* readback) {
   return lovrReadbackIsComplete(readback) ? readback->image : NULL;
 }
@@ -2813,6 +2820,7 @@ Tally* lovrTallyCreate(const TallyInfo* info) {
   lovrCheck(info->count > 0, "Tally count must be greater than zero");
   lovrCheck(info->count <= 4096, "Maximum Tally count is 4096");
   lovrCheck(info->views <= state.limits.renderSize[2], "Tally view count can not exceed the maximum view count");
+  lovrCheck(info->type != TALLY_STAGE || state.features.stageTally, "This GPU does not support the 'stage' Tally type");
   Tally* tally = calloc(1, sizeof(Tally) + gpu_sizeof_tally());
   lovrAssert(tally, "Out of memory");
   tally->ref = 1;
@@ -4864,7 +4872,7 @@ void lovrPassCopyTallyToBuffer(Pass* pass, Tally* tally, Buffer* buffer, uint32_
     lovrTallyResolve(tally, srcIndex, count, buffer->gpu, dstOffset, pass->stream);
     trackBuffer(pass, buffer, GPU_PHASE_SHADER_COMPUTE, GPU_CACHE_STORAGE_WRITE);
   } else {
-    uint32_t stride = tally->info.type == TALLY_STAGE ? 24 : 4;
+    uint32_t stride = tally->info.type == TALLY_STAGE ? 16 : 4;
     gpu_copy_tally_buffer(pass->stream, tally->gpu, buffer->gpu, srcIndex, dstOffset, count, stride);
     trackBuffer(pass, buffer, GPU_PHASE_TRANSFER, GPU_CACHE_TRANSFER_WRITE);
   }
@@ -5005,7 +5013,7 @@ Readback* lovrPassReadTally(Pass* pass, Tally* tally, uint32_t index, uint32_t c
   if (tally->info.type == TALLY_TIMER) {
     lovrTallyResolve(tally, index, count, readback->buffer, 0, pass->stream);
   } else {
-    uint32_t stride = tally->info.type == TALLY_STAGE ? 24 : 4;
+    uint32_t stride = tally->info.type == TALLY_STAGE ? 16 : 4;
     gpu_copy_tally_buffer(pass->stream, tally->gpu, readback->buffer, index, 0, count, stride);
   }
 

src/modules/graphics/graphics.h

@@ -35,6 +35,7 @@ typedef struct {
   bool wireframe;
   bool depthClamp;
   bool indirectDrawFirstInstance;
+  bool stageTally;
   bool float64;
   bool int64;
   bool int16;
@@ -462,6 +463,7 @@ const ReadbackInfo* lovrReadbackGetInfo(Readback* readback);
 bool lovrReadbackIsComplete(Readback* readback);
 bool lovrReadbackWait(Readback* readback);
 void* lovrReadbackGetData(Readback* readback);
+struct Blob* lovrReadbackGetBlob(Readback* readback);
 struct Image* lovrReadbackGetImage(Readback* readback);
 
 // Tally