radv: Implement VK_AMD_shader_info
authorAlex Smith <asmith@feralinteractive.com>
Fri, 27 Oct 2017 13:25:05 +0000 (14:25 +0100)
committerBas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Sat, 28 Oct 2017 22:28:45 +0000 (00:28 +0200)
This allows an app to query shader statistics and get a disassembly of
a shader. RenderDoc git has support for it, so this allows you to view
shader disassembly from a capture.

When this extension is enabled on a device (or when tracing), we now
disable pipeline caching, since we don't get the shader debug info when
we retrieve cached shaders.

v2: Improvements to resource usage reporting
v3: Disassembly string must be null terminated (string_buffer's length
    does not include the terminator)
v4: Fixed LDS reporting. (Bas)

Signed-off-by: Alex Smith <asmith@feralinteractive.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
src/amd/vulkan/radv_device.c
src/amd/vulkan/radv_extensions.py
src/amd/vulkan/radv_pipeline.c
src/amd/vulkan/radv_pipeline_cache.c
src/amd/vulkan/radv_private.h
src/amd/vulkan/radv_shader.c

index d25e9c97ba8c735e4b3659a849ad07070a03d07e..0c2f6fa6312d983c349474809a6b9274a09b86fb 100644 (file)
@@ -944,10 +944,15 @@ VkResult radv_CreateDevice(
        VkResult result;
        struct radv_device *device;
 
+       bool keep_shader_info = false;
+
        for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
                const char *ext_name = pCreateInfo->ppEnabledExtensionNames[i];
                if (!radv_physical_device_extension_supported(physical_device, ext_name))
                        return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
+
+               if (strcmp(ext_name, VK_AMD_SHADER_INFO_EXTENSION_NAME) == 0)
+                       keep_shader_info = true;
        }
 
        /* Check enabled features */
@@ -1041,10 +1046,14 @@ VkResult radv_CreateDevice(
                device->physical_device->rad_info.max_se >= 2;
 
        if (getenv("RADV_TRACE_FILE")) {
+               keep_shader_info = true;
+
                if (!radv_init_trace(device))
                        goto fail;
        }
 
+       device->keep_shader_info = keep_shader_info;
+
        result = radv_device_init_meta(device);
        if (result != VK_SUCCESS)
                goto fail;
index dfeb2880fc25bd1e96f379579a2c494a55da5bd9..eeb679d65a01903fead2d3cf1e8377c856befcf9 100644 (file)
@@ -81,6 +81,7 @@ EXTENSIONS = [
     Extension('VK_EXT_global_priority',                   1, 'device->rad_info.has_ctx_priority'),
     Extension('VK_AMD_draw_indirect_count',               1, True),
     Extension('VK_AMD_rasterization_order',               1, 'device->rad_info.chip_class >= VI && device->rad_info.max_se >= 2'),
+    Extension('VK_AMD_shader_info',                       1, True),
 ]
 
 class VkVersion:
index c25642c9667531ca8b0bc421ce86a43d5d6a3857..c6d9debc7e44d03e2f71f6b720e2a3a162edd3b9 100644 (file)
@@ -1942,7 +1942,7 @@ void radv_create_shaders(struct radv_pipeline *pipeline,
 
        for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
                free(codes[i]);
-               if (modules[i] && !pipeline->device->trace_bo)
+               if (modules[i] && !pipeline->device->keep_shader_info)
                        ralloc_free(nir[i]);
        }
 
index 5dee114749107f697ef449a7d3c5a5df1e7308ad..441e2257d55c67fdf9f0925e74996c2d8ac59b33 100644 (file)
@@ -62,9 +62,11 @@ radv_pipeline_cache_init(struct radv_pipeline_cache *cache,
        cache->hash_table = malloc(byte_size);
 
        /* We don't consider allocation failure fatal, we just start with a 0-sized
-        * cache. */
+        * cache. Disable caching when we want to keep shader debug info, since
+        * we don't get the debug info on cached shaders. */
        if (cache->hash_table == NULL ||
-           (device->instance->debug_flags & RADV_DEBUG_NO_CACHE))
+           (device->instance->debug_flags & RADV_DEBUG_NO_CACHE) ||
+           device->keep_shader_info)
                cache->table_size = 0;
        else
                memset(cache->hash_table, 0, byte_size);
@@ -186,8 +188,11 @@ radv_create_shader_variants_from_pipeline_cache(struct radv_device *device,
        entry = radv_pipeline_cache_search_unlocked(cache, sha1);
 
        if (!entry) {
+               /* Again, don't cache when we want debug info, since this isn't
+                * present in the cache. */
                if (!device->physical_device->disk_cache ||
-                   (device->instance->debug_flags & RADV_DEBUG_NO_CACHE)) {
+                   (device->instance->debug_flags & RADV_DEBUG_NO_CACHE) ||
+                   device->keep_shader_info) {
                        pthread_mutex_unlock(&cache->mutex);
                        return false;
                }
index c19aa070d3808d76d163f28a1b96596c4a7595b5..381afb777f3ebe374ad42528feeb1012177fc0f3 100644 (file)
@@ -563,6 +563,9 @@ struct radv_device {
        struct radeon_winsys_bo                      *trace_bo;
        uint32_t                                     *trace_id_ptr;
 
+       /* Whether to keep shader debug info, for tracing or VK_AMD_shader_info */
+       bool                                         keep_shader_info;
+
        struct radv_physical_device                  *physical_device;
 
        /* Backup in-memory cache to be used if the app doesn't provide one */
index 7f10798fdf435ab60160f5fd51f66e1508eebff7..c9edb28ba2491e99b2de2e670830ec1f47b339b7 100644 (file)
@@ -46,6 +46,8 @@
 #include "util/debug.h"
 #include "ac_exp_param.h"
 
+#include "util/string_buffer.h"
+
 static const struct nir_shader_compiler_options nir_options = {
        .vertex_id_zero_based = true,
        .lower_scmp = true,
@@ -471,7 +473,7 @@ shader_variant_create(struct radv_device *device,
        free(binary.relocs);
        variant->ref_count = 1;
 
-       if (device->trace_bo) {
+       if (device->keep_shader_info) {
                variant->disasm_string = binary.disasm_string;
                if (!gs_copy_shader && !module->nir) {
                        variant->nir = *shaders;
@@ -593,11 +595,20 @@ radv_get_shader_name(struct radv_shader_variant *var, gl_shader_stage stage)
        };
 }
 
-void
-radv_shader_dump_stats(struct radv_device *device,
-                      struct radv_shader_variant *variant,
-                      gl_shader_stage stage,
-                      FILE *file)
+static uint32_t
+get_total_sgprs(struct radv_device *device)
+{
+       if (device->physical_device->rad_info.chip_class >= VI)
+               return 800;
+       else
+               return 512;
+}
+
+static void
+generate_shader_stats(struct radv_device *device,
+                     struct radv_shader_variant *variant,
+                     gl_shader_stage stage,
+                     struct _mesa_string_buffer *buf)
 {
        unsigned lds_increment = device->physical_device->rad_info.chip_class >= CIK ? 512 : 256;
        struct ac_shader_config *conf;
@@ -623,12 +634,8 @@ radv_shader_dump_stats(struct radv_device *device,
                                     lds_increment);
        }
 
-       if (conf->num_sgprs) {
-               if (device->physical_device->rad_info.chip_class >= VI)
-                       max_simd_waves = MIN2(max_simd_waves, 800 / conf->num_sgprs);
-               else
-                       max_simd_waves = MIN2(max_simd_waves, 512 / conf->num_sgprs);
-       }
+       if (conf->num_sgprs)
+               max_simd_waves = MIN2(max_simd_waves, get_total_sgprs(device) / conf->num_sgprs);
 
        if (conf->num_vgprs)
                max_simd_waves = MIN2(max_simd_waves, 256 / conf->num_vgprs);
@@ -639,27 +646,138 @@ radv_shader_dump_stats(struct radv_device *device,
        if (lds_per_wave)
                max_simd_waves = MIN2(max_simd_waves, 16384 / lds_per_wave);
 
+       if (stage == MESA_SHADER_FRAGMENT) {
+               _mesa_string_buffer_printf(buf, "*** SHADER CONFIG ***\n"
+                                          "SPI_PS_INPUT_ADDR = 0x%04x\n"
+                                          "SPI_PS_INPUT_ENA  = 0x%04x\n",
+                                          conf->spi_ps_input_addr, conf->spi_ps_input_ena);
+       }
+
+       _mesa_string_buffer_printf(buf, "*** SHADER STATS ***\n"
+                                  "SGPRS: %d\n"
+                                  "VGPRS: %d\n"
+                                  "Spilled SGPRs: %d\n"
+                                  "Spilled VGPRs: %d\n"
+                                  "Code Size: %d bytes\n"
+                                  "LDS: %d blocks\n"
+                                  "Scratch: %d bytes per wave\n"
+                                  "Max Waves: %d\n"
+                                  "********************\n\n\n",
+                                  conf->num_sgprs, conf->num_vgprs,
+                                  conf->spilled_sgprs, conf->spilled_vgprs, variant->code_size,
+                                  conf->lds_size, conf->scratch_bytes_per_wave,
+                                  max_simd_waves);
+}
+
+void
+radv_shader_dump_stats(struct radv_device *device,
+                      struct radv_shader_variant *variant,
+                      gl_shader_stage stage,
+                      FILE *file)
+{
+       struct _mesa_string_buffer *buf = _mesa_string_buffer_create(NULL, 256);
+
+       generate_shader_stats(device, variant, stage, buf);
+
        fprintf(file, "\n%s:\n", radv_get_shader_name(variant, stage));
+       fprintf(file, buf->buf);
 
-       if (stage == MESA_SHADER_FRAGMENT) {
-               fprintf(file, "*** SHADER CONFIG ***\n"
-                       "SPI_PS_INPUT_ADDR = 0x%04x\n"
-                       "SPI_PS_INPUT_ENA  = 0x%04x\n",
-                       conf->spi_ps_input_addr, conf->spi_ps_input_ena);
+       _mesa_string_buffer_destroy(buf);
+}
+
+VkResult
+radv_GetShaderInfoAMD(VkDevice _device,
+                     VkPipeline _pipeline,
+                     VkShaderStageFlagBits shaderStage,
+                     VkShaderInfoTypeAMD infoType,
+                     size_t* pInfoSize,
+                     void* pInfo)
+{
+       RADV_FROM_HANDLE(radv_device, device, _device);
+       RADV_FROM_HANDLE(radv_pipeline, pipeline, _pipeline);
+       gl_shader_stage stage = vk_to_mesa_shader_stage(shaderStage);
+       struct radv_shader_variant *variant = pipeline->shaders[stage];
+       struct _mesa_string_buffer *buf;
+       VkResult result = VK_SUCCESS;
+
+       /* Spec doesn't indicate what to do if the stage is invalid, so just
+        * return no info for this. */
+       if (!variant)
+               return VK_ERROR_FEATURE_NOT_PRESENT;
+
+       switch (infoType) {
+       case VK_SHADER_INFO_TYPE_STATISTICS_AMD:
+               if (!pInfo) {
+                       *pInfoSize = sizeof(VkShaderStatisticsInfoAMD);
+               } else {
+                       unsigned lds_multiplier = device->physical_device->rad_info.chip_class >= CIK ? 512 : 256;
+                       struct ac_shader_config *conf = &variant->config;
+
+                       VkShaderStatisticsInfoAMD statistics = {};
+                       statistics.shaderStageMask = shaderStage;
+                       statistics.numPhysicalVgprs = 256;
+                       statistics.numPhysicalSgprs = get_total_sgprs(device);
+                       statistics.numAvailableSgprs = statistics.numPhysicalSgprs;
+
+                       if (stage == MESA_SHADER_COMPUTE) {
+                               unsigned *local_size = variant->nir->info.cs.local_size;
+                               unsigned workgroup_size = local_size[0] * local_size[1] * local_size[2];
+
+                               statistics.numAvailableVgprs = statistics.numPhysicalVgprs /
+                                                              ceil(workgroup_size / statistics.numPhysicalVgprs);
+
+                               statistics.computeWorkGroupSize[0] = local_size[0];
+                               statistics.computeWorkGroupSize[1] = local_size[1];
+                               statistics.computeWorkGroupSize[2] = local_size[2];
+                       } else {
+                               statistics.numAvailableVgprs = statistics.numPhysicalVgprs;
+                       }
+
+                       statistics.resourceUsage.numUsedVgprs = conf->num_vgprs;
+                       statistics.resourceUsage.numUsedSgprs = conf->num_sgprs;
+                       statistics.resourceUsage.ldsSizePerLocalWorkGroup = 32768;
+                       statistics.resourceUsage.ldsUsageSizeInBytes = conf->lds_size * lds_multiplier;
+                       statistics.resourceUsage.scratchMemUsageInBytes = conf->scratch_bytes_per_wave;
+
+                       size_t size = *pInfoSize;
+                       *pInfoSize = sizeof(statistics);
+
+                       memcpy(pInfo, &statistics, MIN2(size, *pInfoSize));
+
+                       if (size < *pInfoSize)
+                               result = VK_INCOMPLETE;
+               }
+
+               break;
+       case VK_SHADER_INFO_TYPE_DISASSEMBLY_AMD:
+               buf = _mesa_string_buffer_create(NULL, 1024);
+
+               _mesa_string_buffer_printf(buf, "%s:\n", radv_get_shader_name(variant, stage));
+               _mesa_string_buffer_printf(buf, "%s\n\n", variant->disasm_string);
+               generate_shader_stats(device, variant, stage, buf);
+
+               /* Need to include the null terminator. */
+               size_t length = buf->length + 1;
+
+               if (!pInfo) {
+                       *pInfoSize = length;
+               } else {
+                       size_t size = *pInfoSize;
+                       *pInfoSize = length;
+
+                       memcpy(pInfo, buf->buf, MIN2(size, length));
+
+                       if (size < length)
+                               result = VK_INCOMPLETE;
+               }
+
+               _mesa_string_buffer_destroy(buf);
+               break;
+       default:
+               /* VK_SHADER_INFO_TYPE_BINARY_AMD unimplemented for now. */
+               result = VK_ERROR_FEATURE_NOT_PRESENT;
+               break;
        }
 
-       fprintf(file, "*** SHADER STATS ***\n"
-               "SGPRS: %d\n"
-               "VGPRS: %d\n"
-               "Spilled SGPRs: %d\n"
-               "Spilled VGPRs: %d\n"
-               "Code Size: %d bytes\n"
-               "LDS: %d blocks\n"
-               "Scratch: %d bytes per wave\n"
-               "Max Waves: %d\n"
-               "********************\n\n\n",
-               conf->num_sgprs, conf->num_vgprs,
-               conf->spilled_sgprs, conf->spilled_vgprs, variant->code_size,
-               conf->lds_size, conf->scratch_bytes_per_wave,
-               max_simd_waves);
+       return result;
 }