#include <stdio.h>
#include <sys/utsname.h>
+#include "util/mesa-sha1.h"
#include "sid.h"
-#include "gfx9d.h"
#include "ac_debug.h"
#include "radv_debug.h"
#include "radv_shader.h"
#define TRACE_BO_SIZE 4096
+#define TMA_BO_SIZE 4096
#define COLOR_RESET "\033[0m"
#define COLOR_RED "\033[31m"
device->trace_bo = ws->buffer_create(ws, TRACE_BO_SIZE, 8,
RADEON_DOMAIN_VRAM,
RADEON_FLAG_CPU_ACCESS|
- RADEON_FLAG_NO_INTERPROCESS_SHARING);
+ RADEON_FLAG_NO_INTERPROCESS_SHARING |
+ RADEON_FLAG_ZERO_VRAM,
+ RADV_BO_PRIORITY_UPLOAD_BUFFER);
if (!device->trace_bo)
return false;
if (!device->trace_id_ptr)
return false;
- memset(device->trace_id_ptr, 0, TRACE_BO_SIZE);
-
ac_vm_fault_occured(device->physical_device->rad_info.chip_class,
&device->dmesg_timestamp, NULL);
}
static void
-radv_dump_trace(struct radv_device *device, struct radeon_winsys_cs *cs)
+radv_dump_trace(struct radv_device *device, struct radeon_cmdbuf *cs)
{
const char *filename = getenv("RADV_TRACE_FILE");
FILE *f = fopen(filename, "w");
{
struct radeon_info *info = &device->physical_device->rad_info;
- if (info->drm_major == 2 && info->drm_minor < 42)
- return; /* no radeon support */
-
fprintf(f, "Memory-mapped registers:\n");
radv_dump_mmapped_reg(device, f, R_008010_GRBM_STATUS);
/* No other registers can be read on DRM < 3.1.0. */
- if (info->drm_major < 3 || info->drm_minor < 1) {
+ if (info->drm_minor < 1) {
fprintf(f, "\n");
return;
}
radv_dump_mmapped_reg(device, f, R_00803C_GRBM_STATUS_SE3);
radv_dump_mmapped_reg(device, f, R_00D034_SDMA0_STATUS_REG);
radv_dump_mmapped_reg(device, f, R_00D834_SDMA1_STATUS_REG);
- if (info->chip_class <= VI) {
+ if (info->chip_class <= GFX8) {
radv_dump_mmapped_reg(device, f, R_000E50_SRBM_STATUS);
radv_dump_mmapped_reg(device, f, R_000E4C_SRBM_STATUS2);
radv_dump_mmapped_reg(device, f, R_000E54_SRBM_STATUS3);
fprintf(f, "\n");
}
-static const char *
-radv_get_descriptor_name(enum VkDescriptorType type)
-{
- switch (type) {
- case VK_DESCRIPTOR_TYPE_SAMPLER:
- return "SAMPLER";
- case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
- return "COMBINED_IMAGE_SAMPLER";
- case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
- return "SAMPLED_IMAGE";
- case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
- return "STORAGE_IMAGE";
- case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
- return "UNIFORM_TEXEL_BUFFER";
- case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
- return "STORAGE_TEXEL_BUFFER";
- case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
- return "UNIFORM_BUFFER";
- case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
- return "STORAGE_BUFFER";
- case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
- return "UNIFORM_BUFFER_DYNAMIC";
- case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
- return "STORAGE_BUFFER_DYNAMIC";
- case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
- return "INPUT_ATTACHMENT";
- default:
- return "UNKNOWN";
- }
-}
-
static void
radv_dump_buffer_descriptor(enum chip_class chip_class, const uint32_t *desc,
FILE *f)
radv_dump_image_descriptor(enum chip_class chip_class, const uint32_t *desc,
FILE *f)
{
+ unsigned sq_img_rsrc_word0 = chip_class >= GFX10 ? R_00A000_SQ_IMG_RSRC_WORD0
+ : R_008F10_SQ_IMG_RSRC_WORD0;
+
fprintf(f, COLOR_CYAN " Image:" COLOR_RESET "\n");
for (unsigned j = 0; j < 8; j++)
- ac_dump_reg(f, chip_class, R_008F10_SQ_IMG_RSRC_WORD0 + j * 4,
+ ac_dump_reg(f, chip_class, sq_img_rsrc_word0 + j * 4,
desc[j], 0xffffffff);
fprintf(f, COLOR_CYAN " FMASK:" COLOR_RESET "\n");
for (unsigned j = 0; j < 8; j++)
- ac_dump_reg(f, chip_class, R_008F10_SQ_IMG_RSRC_WORD0 + j * 4,
+ ac_dump_reg(f, chip_class, sq_img_rsrc_word0 + j * 4,
desc[8 + j], 0xffffffff);
}
}
static void
-radv_dump_descriptor_set(enum chip_class chip_class,
+radv_dump_descriptor_set(struct radv_device *device,
struct radv_descriptor_set *set, unsigned id, FILE *f)
{
+ enum chip_class chip_class = device->physical_device->rad_info.chip_class;
const struct radv_descriptor_set_layout *layout;
int i;
return;
layout = set->layout;
- fprintf(f, "** descriptor set (%d) **\n", id);
- fprintf(f, "va: 0x%"PRIx64"\n", set->va);
- fprintf(f, "size: %d\n", set->size);
- fprintf(f, "mapped_ptr:\n");
-
- for (i = 0; i < set->size / 4; i++) {
- fprintf(f, "\t[0x%x] = 0x%08x\n", i, set->mapped_ptr[i]);
- }
- fprintf(f, "\n");
-
- fprintf(f, "\t*** layout ***\n");
- fprintf(f, "\tbinding_count: %d\n", layout->binding_count);
- fprintf(f, "\tsize: %d\n", layout->size);
- fprintf(f, "\tshader_stages: %x\n", layout->shader_stages);
- fprintf(f, "\tdynamic_shader_stages: %x\n",
- layout->dynamic_shader_stages);
- fprintf(f, "\tbuffer_count: %d\n", layout->buffer_count);
- fprintf(f, "\tdynamic_offset_count: %d\n",
- layout->dynamic_offset_count);
- fprintf(f, "\n");
-
for (i = 0; i < set->layout->binding_count; i++) {
uint32_t *desc =
set->mapped_ptr + layout->binding[i].offset / 4;
- fprintf(f, "\t\t**** binding layout (%d) ****\n", i);
- fprintf(f, "\t\ttype: %s\n",
- radv_get_descriptor_name(layout->binding[i].type));
- fprintf(f, "\t\tarray_size: %d\n",
- layout->binding[i].array_size);
- fprintf(f, "\t\toffset: %d\n",
- layout->binding[i].offset);
- fprintf(f, "\t\tbuffer_offset: %d\n",
- layout->binding[i].buffer_offset);
- fprintf(f, "\t\tdynamic_offset_offset: %d\n",
- layout->binding[i].dynamic_offset_offset);
- fprintf(f, "\t\tdynamic_offset_count: %d\n",
- layout->binding[i].dynamic_offset_count);
- fprintf(f, "\t\tsize: %d\n",
- layout->binding[i].size);
- fprintf(f, "\t\timmutable_samplers_offset: %d\n",
- layout->binding[i].immutable_samplers_offset);
- fprintf(f, "\t\timmutable_samplers_equal: %d\n",
- layout->binding[i].immutable_samplers_equal);
- fprintf(f, "\n");
-
switch (layout->binding[i].type) {
case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
}
static void
-radv_dump_descriptors(struct radv_pipeline *pipeline, FILE *f)
+radv_dump_descriptors(struct radv_device *device, FILE *f)
{
- struct radv_device *device = pipeline->device;
- enum chip_class chip_class = device->physical_device->rad_info.chip_class;
uint64_t *ptr = (uint64_t *)device->trace_id_ptr;
int i;
- fprintf(f, "List of descriptors:\n");
+ fprintf(f, "Descriptors:\n");
for (i = 0; i < MAX_SETS; i++) {
struct radv_descriptor_set *set =
- (struct radv_descriptor_set *)ptr[i + 3];
+ *(struct radv_descriptor_set **)(ptr + i + 3);
- radv_dump_descriptor_set(chip_class, set, i, f);
+ radv_dump_descriptor_set(device, set, i, f);
}
}
struct radv_shader_inst *inst = &instructions[*num];
unsigned len = next - disasm;
+ if (!memchr(disasm, ';', len)) {
+ /* Ignore everything that is not an instruction. */
+ disasm = next + 1;
+ continue;
+ }
+
assert(len < ARRAY_SIZE(inst->text));
memcpy(inst->text, disasm, len);
inst->text[len] = 0;
}
static void
-radv_dump_annotated_shader(struct radv_pipeline *pipeline,
- struct radv_shader_variant *shader,
- gl_shader_stage stage,
- struct ac_wave_info *waves, unsigned num_waves,
- FILE *f)
+radv_dump_annotated_shader(struct radv_shader_variant *shader,
+ gl_shader_stage stage, struct ac_wave_info *waves,
+ unsigned num_waves, FILE *f)
{
uint64_t start_addr, end_addr;
unsigned i;
start_addr, &num_inst, instructions);
fprintf(f, COLOR_YELLOW "%s - annotated disassembly:" COLOR_RESET "\n",
- radv_get_shader_name(shader, stage));
+ radv_get_shader_name(&shader->info, stage));
/* Print instructions with annotations. */
for (i = 0; i < num_inst; i++) {
static void
radv_dump_annotated_shaders(struct radv_pipeline *pipeline,
- struct radv_shader_variant *compute_shader,
- FILE *f)
+ VkShaderStageFlagBits active_stages, FILE *f)
{
struct ac_wave_info waves[AC_MAX_WAVES_PER_CHIP];
- unsigned num_waves = ac_get_wave_info(waves);
- unsigned mask;
+ enum chip_class chip_class = pipeline->device->physical_device->rad_info.chip_class;
+ unsigned num_waves = ac_get_wave_info(chip_class, waves);
fprintf(f, COLOR_CYAN "The number of active waves = %u" COLOR_RESET
"\n\n", num_waves);
/* Dump annotated active graphics shaders. */
- mask = pipeline->active_stages;
- while (mask) {
- int stage = u_bit_scan(&mask);
+ while (active_stages) {
+ int stage = u_bit_scan(&active_stages);
- radv_dump_annotated_shader(pipeline, pipeline->shaders[stage],
+ radv_dump_annotated_shader(pipeline->shaders[stage],
stage, waves, num_waves, f);
}
- radv_dump_annotated_shader(pipeline, compute_shader,
- MESA_SHADER_COMPUTE, waves, num_waves, f);
-
/* Print waves executing shaders that are not currently bound. */
unsigned i;
bool found = false;
if (!shader)
return;
- fprintf(f, "%s:\n\n", radv_get_shader_name(shader, stage));
+ fprintf(f, "%s:\n\n", radv_get_shader_name(&shader->info, stage));
if (shader->spirv) {
- fprintf(f, "SPIRV:\n");
+ unsigned char sha1[21];
+ char sha1buf[41];
+
+ _mesa_sha1_compute(shader->spirv, shader->spirv_size, sha1);
+ _mesa_sha1_format(sha1buf, sha1);
+
+ fprintf(f, "SPIRV (sha1: %s):\n", sha1buf);
radv_print_spirv(shader->spirv, shader->spirv_size, f);
}
- if (shader->nir) {
- fprintf(f, "NIR:\n");
- nir_print_shader(shader->nir, f);
+ if (shader->nir_string) {
+ fprintf(f, "NIR:\n%s\n", shader->nir_string);
}
+ fprintf(f, "%s IR:\n%s\n",
+ pipeline->device->physical_device->use_llvm ? "LLVM" : "ACO",
+ shader->ir_string);
fprintf(f, "DISASM:\n%s\n", shader->disasm_string);
radv_shader_dump_stats(pipeline->device, shader, stage, f);
static void
radv_dump_shaders(struct radv_pipeline *pipeline,
- struct radv_shader_variant *compute_shader, FILE *f)
+ VkShaderStageFlagBits active_stages, FILE *f)
{
- unsigned mask;
-
/* Dump active graphics shaders. */
- mask = pipeline->active_stages;
- while (mask) {
- int stage = u_bit_scan(&mask);
+ while (active_stages) {
+ int stage = u_bit_scan(&active_stages);
radv_dump_shader(pipeline, pipeline->shaders[stage], stage, f);
}
+}
- radv_dump_shader(pipeline, compute_shader, MESA_SHADER_COMPUTE, f);
+static void
+radv_dump_pipeline_state(struct radv_pipeline *pipeline,
+ VkShaderStageFlagBits active_stages, FILE *f)
+{
+ radv_dump_shaders(pipeline, active_stages, f);
+ radv_dump_annotated_shaders(pipeline, active_stages, f);
}
static void
-radv_dump_graphics_state(struct radv_pipeline *graphics_pipeline,
+radv_dump_graphics_state(struct radv_device *device,
+ struct radv_pipeline *graphics_pipeline,
struct radv_pipeline *compute_pipeline, FILE *f)
{
- struct radv_shader_variant *compute_shader =
- compute_pipeline ? compute_pipeline->shaders[MESA_SHADER_COMPUTE] : NULL;
+ VkShaderStageFlagBits active_stages;
- if (!graphics_pipeline)
- return;
+ if (graphics_pipeline) {
+ active_stages = graphics_pipeline->active_stages;
+ radv_dump_pipeline_state(graphics_pipeline, active_stages, f);
+ }
- radv_dump_shaders(graphics_pipeline, compute_shader, f);
- radv_dump_annotated_shaders(graphics_pipeline, compute_shader, f);
- radv_dump_descriptors(graphics_pipeline, f);
+ if (compute_pipeline) {
+ active_stages = VK_SHADER_STAGE_COMPUTE_BIT;
+ radv_dump_pipeline_state(compute_pipeline, active_stages, f);
+ }
+
+ radv_dump_descriptors(device, f);
}
static void
-radv_dump_compute_state(struct radv_pipeline *compute_pipeline, FILE *f)
+radv_dump_compute_state(struct radv_device *device,
+ struct radv_pipeline *compute_pipeline, FILE *f)
{
+ VkShaderStageFlagBits active_stages = VK_SHADER_STAGE_COMPUTE_BIT;
+
if (!compute_pipeline)
return;
- radv_dump_shaders(compute_pipeline,
- compute_pipeline->shaders[MESA_SHADER_COMPUTE], f);
- radv_dump_annotated_shaders(compute_pipeline,
- compute_pipeline->shaders[MESA_SHADER_COMPUTE],
- f);
- radv_dump_descriptors(compute_pipeline, f);
+ radv_dump_pipeline_state(compute_pipeline, active_stages, f);
+ radv_dump_descriptors(device, f);
}
static struct radv_pipeline *
{
uint64_t *ptr = (uint64_t *)device->trace_id_ptr;
- return (struct radv_pipeline *)ptr[1];
+ return *(struct radv_pipeline **)(ptr + 1);
}
static struct radv_pipeline *
{
uint64_t *ptr = (uint64_t *)device->trace_id_ptr;
- return (struct radv_pipeline *)ptr[2];
+ return *(struct radv_pipeline **)(ptr + 2);
}
static void
radv_dump_device_name(struct radv_device *device, FILE *f)
{
struct radeon_info *info = &device->physical_device->rad_info;
- char llvm_string[32] = {}, kernel_version[128] = {};
+ char kernel_version[128] = {};
struct utsname uname_data;
const char *chip_name;
snprintf(kernel_version, sizeof(kernel_version),
" / %s", uname_data.release);
- if (HAVE_LLVM > 0) {
- snprintf(llvm_string, sizeof(llvm_string),
- ", LLVM %i.%i.%i", (HAVE_LLVM >> 8) & 0xff,
- HAVE_LLVM & 0xff, MESA_LLVM_VERSION_PATCH);
- }
-
- fprintf(f, "Device name: %s (%s DRM %i.%i.%i%s%s)\n\n",
+ fprintf(f, "Device name: %s (%s / DRM %i.%i.%i%s)\n\n",
chip_name, device->physical_device->name,
info->drm_major, info->drm_minor, info->drm_patchlevel,
- kernel_version, llvm_string);
+ kernel_version);
}
static bool
}
void
-radv_check_gpu_hangs(struct radv_queue *queue, struct radeon_winsys_cs *cs)
+radv_check_gpu_hangs(struct radv_queue *queue, struct radeon_cmdbuf *cs)
{
struct radv_pipeline *graphics_pipeline, *compute_pipeline;
struct radv_device *device = queue->device;
graphics_pipeline = radv_get_saved_graphics_pipeline(device);
compute_pipeline = radv_get_saved_compute_pipeline(device);
+ radv_dump_trace(queue->device, cs);
+
fprintf(stderr, "GPU hang report:\n\n");
radv_dump_device_name(device, stderr);
switch (ring) {
case RING_GFX:
- radv_dump_graphics_state(graphics_pipeline, compute_pipeline,
+ fprintf(stderr, "RING_GFX:\n");
+ radv_dump_graphics_state(queue->device,
+ graphics_pipeline, compute_pipeline,
stderr);
break;
case RING_COMPUTE:
- radv_dump_compute_state(compute_pipeline, stderr);
+ fprintf(stderr, "RING_COMPUTE:\n");
+ radv_dump_compute_state(queue->device,
+ compute_pipeline, stderr);
break;
default:
assert(0);
break;
}
- radv_dump_trace(queue->device, cs);
abort();
}
void
-radv_print_spirv(uint32_t *data, uint32_t size, FILE *fp)
+radv_print_spirv(const char *data, uint32_t size, FILE *fp)
{
char path[] = "/tmp/fileXXXXXX";
char line[2048], command[128];
close(fd);
unlink(path);
}
+
+bool
+radv_trap_handler_init(struct radv_device *device)
+{
+ struct radeon_winsys *ws = device->ws;
+
+ /* Create the trap handler shader and upload it like other shaders. */
+ device->trap_handler_shader = radv_create_trap_handler_shader(device);
+ if (!device->trap_handler_shader) {
+ fprintf(stderr, "radv: failed to create the trap handler shader.\n");
+ return false;
+ }
+
+ device->tma_bo = ws->buffer_create(ws, TMA_BO_SIZE, 8,
+ RADEON_DOMAIN_VRAM,
+ RADEON_FLAG_CPU_ACCESS |
+ RADEON_FLAG_NO_INTERPROCESS_SHARING |
+ RADEON_FLAG_ZERO_VRAM,
+ RADV_BO_PRIORITY_SCRATCH);
+ if (!device->tma_bo)
+ return false;
+
+ device->tma_ptr = ws->buffer_map(device->tma_bo);
+ if (!device->tma_ptr)
+ return false;
+
+ /* Upload a buffer descriptor to store various info from the trap. */
+ uint64_t tma_va = radv_buffer_get_va(device->tma_bo) + 16;
+ uint32_t desc[4];
+
+ desc[0] = tma_va;
+ desc[1] = S_008F04_BASE_ADDRESS_HI(tma_va >> 32);
+ desc[2] = TMA_BO_SIZE;
+ desc[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
+ S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
+ S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
+ S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
+ S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
+
+ memcpy(device->tma_ptr, desc, sizeof(desc));
+
+ return true;
+}
+
+void
+radv_trap_handler_finish(struct radv_device *device)
+{
+ struct radeon_winsys *ws = device->ws;
+
+ if (unlikely(device->trap_handler_shader))
+ radv_shader_variant_destroy(device, device->trap_handler_shader);
+
+ if (unlikely(device->tma_bo))
+ ws->buffer_destroy(device->tma_bo);
+}
projects / mesa.git / blobdiff