/*
* Copyright 2015 Advanced Micro Devices, Inc.
+ * All Rights Reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* 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.
- *
- * Authors:
- * Marek Olšák <maraeo@gmail.com>
*/
#include "si_pipe.h"
#include "sid.h"
#include "gfx9d.h"
#include "sid_tables.h"
-#include "ddebug/dd_util.h"
+#include "driver_ddebug/dd_util.h"
+#include "util/u_dump.h"
+#include "util/u_log.h"
#include "util/u_memory.h"
+#include "util/u_string.h"
#include "ac_debug.h"
+static void si_dump_bo_list(struct si_context *sctx,
+ const struct radeon_saved_cs *saved, FILE *f);
+
DEBUG_GET_ONCE_OPTION(replace_shaders, "RADEON_REPLACE_SHADERS", NULL)
+/**
+ * Store a linearized copy of all chunks of \p cs together with the buffer
+ * list in \p saved.
+ */
+void si_save_cs(struct radeon_winsys *ws, struct radeon_cmdbuf *cs,
+ struct radeon_saved_cs *saved, bool get_buffer_list)
+{
+ uint32_t *buf;
+ unsigned i;
+
+ /* Save the IB chunks. */
+ saved->num_dw = cs->prev_dw + cs->current.cdw;
+ saved->ib = MALLOC(4 * saved->num_dw);
+ if (!saved->ib)
+ goto oom;
+
+ buf = saved->ib;
+ for (i = 0; i < cs->num_prev; ++i) {
+ memcpy(buf, cs->prev[i].buf, cs->prev[i].cdw * 4);
+ buf += cs->prev[i].cdw;
+ }
+ memcpy(buf, cs->current.buf, cs->current.cdw * 4);
+
+ if (!get_buffer_list)
+ return;
+
+ /* Save the buffer list. */
+ saved->bo_count = ws->cs_get_buffer_list(cs, NULL);
+ saved->bo_list = CALLOC(saved->bo_count,
+ sizeof(saved->bo_list[0]));
+ if (!saved->bo_list) {
+ FREE(saved->ib);
+ goto oom;
+ }
+ ws->cs_get_buffer_list(cs, saved->bo_list);
+
+ return;
+
+oom:
+ fprintf(stderr, "%s: out of memory\n", __func__);
+ memset(saved, 0, sizeof(*saved));
+}
+
+void si_clear_saved_cs(struct radeon_saved_cs *saved)
+{
+ FREE(saved->ib);
+ FREE(saved->bo_list);
+
+ memset(saved, 0, sizeof(*saved));
+}
+
+void si_destroy_saved_cs(struct si_saved_cs *scs)
+{
+ si_clear_saved_cs(&scs->gfx);
+ r600_resource_reference(&scs->trace_buf, NULL);
+ free(scs);
+}
+
static void si_dump_shader(struct si_screen *sscreen,
enum pipe_shader_type processor,
const struct si_shader *shader, FILE *f)
si_shader_dump(sscreen, shader, NULL, processor, f, false);
}
-static void si_dump_gfx_shader(struct si_screen *sscreen,
- const struct si_shader_ctx_state *state, FILE *f)
+struct si_log_chunk_shader {
+ /* The shader destroy code assumes a current context for unlinking of
+ * PM4 packets etc.
+ *
+ * While we should be able to destroy shaders without a context, doing
+ * so would happen only very rarely and be therefore likely to fail
+ * just when you're trying to debug something. Let's just remember the
+ * current context in the chunk.
+ */
+ struct si_context *ctx;
+ struct si_shader *shader;
+ enum pipe_shader_type processor;
+
+ /* For keep-alive reference counts */
+ struct si_shader_selector *sel;
+ struct si_compute *program;
+};
+
+static void
+si_log_chunk_shader_destroy(void *data)
+{
+ struct si_log_chunk_shader *chunk = data;
+ si_shader_selector_reference(chunk->ctx, &chunk->sel, NULL);
+ si_compute_reference(&chunk->program, NULL);
+ FREE(chunk);
+}
+
+static void
+si_log_chunk_shader_print(void *data, FILE *f)
+{
+ struct si_log_chunk_shader *chunk = data;
+ struct si_screen *sscreen = chunk->ctx->screen;
+ si_dump_shader(sscreen, chunk->processor,
+ chunk->shader, f);
+}
+
+static struct u_log_chunk_type si_log_chunk_type_shader = {
+ .destroy = si_log_chunk_shader_destroy,
+ .print = si_log_chunk_shader_print,
+};
+
+static void si_dump_gfx_shader(struct si_context *ctx,
+ const struct si_shader_ctx_state *state,
+ struct u_log_context *log)
{
- const struct si_shader *current = state->current;
+ struct si_shader *current = state->current;
if (!state->cso || !current)
return;
- si_dump_shader(sscreen, state->cso->info.processor, current, f);
+ struct si_log_chunk_shader *chunk = CALLOC_STRUCT(si_log_chunk_shader);
+ chunk->ctx = ctx;
+ chunk->processor = state->cso->info.processor;
+ chunk->shader = current;
+ si_shader_selector_reference(ctx, &chunk->sel, current->selector);
+ u_log_chunk(log, &si_log_chunk_type_shader, chunk);
}
-static void si_dump_compute_shader(struct si_screen *sscreen,
- const struct si_cs_shader_state *state, FILE *f)
+static void si_dump_compute_shader(struct si_context *ctx,
+ struct u_log_context *log)
{
- if (!state->program || state->program != state->emitted_program)
+ const struct si_cs_shader_state *state = &ctx->cs_shader_state;
+
+ if (!state->program)
return;
- si_dump_shader(sscreen, PIPE_SHADER_COMPUTE, &state->program->shader, f);
+ struct si_log_chunk_shader *chunk = CALLOC_STRUCT(si_log_chunk_shader);
+ chunk->ctx = ctx;
+ chunk->processor = PIPE_SHADER_COMPUTE;
+ chunk->shader = &state->program->shader;
+ si_compute_reference(&chunk->program, state->program);
+ u_log_chunk(log, &si_log_chunk_type_shader, chunk);
}
/**
static void si_dump_mmapped_reg(struct si_context *sctx, FILE *f,
unsigned offset)
{
- struct radeon_winsys *ws = sctx->b.ws;
+ struct radeon_winsys *ws = sctx->ws;
uint32_t value;
if (ws->read_registers(ws, offset, 1, &value))
- ac_dump_reg(f, offset, value, ~0);
+ ac_dump_reg(f, sctx->chip_class, offset, value, ~0);
}
static void si_dump_debug_registers(struct si_context *sctx, FILE *f)
{
- if (sctx->screen->b.info.drm_major == 2 &&
- sctx->screen->b.info.drm_minor < 42)
- return; /* no radeon support */
+ if (!sctx->screen->info.has_read_registers_query)
+ return;
fprintf(f, "Memory-mapped registers:\n");
si_dump_mmapped_reg(sctx, f, R_008010_GRBM_STATUS);
/* No other registers can be read on DRM < 3.1.0. */
- if (sctx->screen->b.info.drm_major < 3 ||
- sctx->screen->b.info.drm_minor < 1) {
+ if (sctx->screen->info.drm_major < 3 ||
+ sctx->screen->info.drm_minor < 1) {
fprintf(f, "\n");
return;
}
si_dump_mmapped_reg(sctx, f, R_00803C_GRBM_STATUS_SE3);
si_dump_mmapped_reg(sctx, f, R_00D034_SDMA0_STATUS_REG);
si_dump_mmapped_reg(sctx, f, R_00D834_SDMA1_STATUS_REG);
- if (sctx->b.chip_class <= VI) {
+ if (sctx->chip_class <= VI) {
si_dump_mmapped_reg(sctx, f, R_000E50_SRBM_STATUS);
si_dump_mmapped_reg(sctx, f, R_000E4C_SRBM_STATUS2);
si_dump_mmapped_reg(sctx, f, R_000E54_SRBM_STATUS3);
fprintf(f, "\n");
}
-static void si_dump_last_ib(struct si_context *sctx, FILE *f)
+struct si_log_chunk_cs {
+ struct si_context *ctx;
+ struct si_saved_cs *cs;
+ bool dump_bo_list;
+ unsigned gfx_begin, gfx_end;
+};
+
+static void si_log_chunk_type_cs_destroy(void *data)
+{
+ struct si_log_chunk_cs *chunk = data;
+ si_saved_cs_reference(&chunk->cs, NULL);
+ free(chunk);
+}
+
+static void si_parse_current_ib(FILE *f, struct radeon_cmdbuf *cs,
+ unsigned begin, unsigned end,
+ int *last_trace_id, unsigned trace_id_count,
+ const char *name, enum chip_class chip_class)
{
+ unsigned orig_end = end;
+
+ assert(begin <= end);
+
+ fprintf(f, "------------------ %s begin (dw = %u) ------------------\n",
+ name, begin);
+
+ for (unsigned prev_idx = 0; prev_idx < cs->num_prev; ++prev_idx) {
+ struct radeon_cmdbuf_chunk *chunk = &cs->prev[prev_idx];
+
+ if (begin < chunk->cdw) {
+ ac_parse_ib_chunk(f, chunk->buf + begin,
+ MIN2(end, chunk->cdw) - begin,
+ last_trace_id, trace_id_count,
+ chip_class, NULL, NULL);
+ }
+
+ if (end <= chunk->cdw)
+ return;
+
+ if (begin < chunk->cdw)
+ fprintf(f, "\n---------- Next %s Chunk ----------\n\n",
+ name);
+
+ begin -= MIN2(begin, chunk->cdw);
+ end -= chunk->cdw;
+ }
+
+ assert(end <= cs->current.cdw);
+
+ ac_parse_ib_chunk(f, cs->current.buf + begin, end - begin, last_trace_id,
+ trace_id_count, chip_class, NULL, NULL);
+
+ fprintf(f, "------------------- %s end (dw = %u) -------------------\n\n",
+ name, orig_end);
+}
+
+static void si_log_chunk_type_cs_print(void *data, FILE *f)
+{
+ struct si_log_chunk_cs *chunk = data;
+ struct si_context *ctx = chunk->ctx;
+ struct si_saved_cs *scs = chunk->cs;
int last_trace_id = -1;
- if (!sctx->last_gfx.ib)
- return;
+ /* We are expecting that the ddebug pipe has already
+ * waited for the context, so this buffer should be idle.
+ * If the GPU is hung, there is no point in waiting for it.
+ */
+ uint32_t *map = ctx->ws->buffer_map(scs->trace_buf->buf,
+ NULL,
+ PIPE_TRANSFER_UNSYNCHRONIZED |
+ PIPE_TRANSFER_READ);
+ if (map)
+ last_trace_id = map[0];
+
+ if (chunk->gfx_end != chunk->gfx_begin) {
+ if (chunk->gfx_begin == 0) {
+ if (ctx->init_config)
+ ac_parse_ib(f, ctx->init_config->pm4, ctx->init_config->ndw,
+ NULL, 0, "IB2: Init config", ctx->chip_class,
+ NULL, NULL);
+
+ if (ctx->init_config_gs_rings)
+ ac_parse_ib(f, ctx->init_config_gs_rings->pm4,
+ ctx->init_config_gs_rings->ndw,
+ NULL, 0, "IB2: Init GS rings", ctx->chip_class,
+ NULL, NULL);
+ }
+
+ if (scs->flushed) {
+ ac_parse_ib(f, scs->gfx.ib + chunk->gfx_begin,
+ chunk->gfx_end - chunk->gfx_begin,
+ &last_trace_id, map ? 1 : 0, "IB", ctx->chip_class,
+ NULL, NULL);
+ } else {
+ si_parse_current_ib(f, ctx->gfx_cs, chunk->gfx_begin,
+ chunk->gfx_end, &last_trace_id, map ? 1 : 0,
+ "IB", ctx->chip_class);
+ }
+ }
- if (sctx->last_trace_buf) {
- /* We are expecting that the ddebug pipe has already
- * waited for the context, so this buffer should be idle.
- * If the GPU is hung, there is no point in waiting for it.
- */
- uint32_t *map = sctx->b.ws->buffer_map(sctx->last_trace_buf->buf,
- NULL,
- PIPE_TRANSFER_UNSYNCHRONIZED |
- PIPE_TRANSFER_READ);
- if (map)
- last_trace_id = *map;
+ if (chunk->dump_bo_list) {
+ fprintf(f, "Flushing. Time: ");
+ util_dump_ns(f, scs->time_flush);
+ fprintf(f, "\n\n");
+ si_dump_bo_list(ctx, &scs->gfx, f);
}
+}
+
+static const struct u_log_chunk_type si_log_chunk_type_cs = {
+ .destroy = si_log_chunk_type_cs_destroy,
+ .print = si_log_chunk_type_cs_print,
+};
+
+static void si_log_cs(struct si_context *ctx, struct u_log_context *log,
+ bool dump_bo_list)
+{
+ assert(ctx->current_saved_cs);
+
+ struct si_saved_cs *scs = ctx->current_saved_cs;
+ unsigned gfx_cur = ctx->gfx_cs->prev_dw + ctx->gfx_cs->current.cdw;
- if (sctx->init_config)
- ac_parse_ib(f, sctx->init_config->pm4, sctx->init_config->ndw,
- -1, "IB2: Init config", sctx->b.chip_class,
- NULL, NULL);
+ if (!dump_bo_list &&
+ gfx_cur == scs->gfx_last_dw)
+ return;
+
+ struct si_log_chunk_cs *chunk = calloc(1, sizeof(*chunk));
+
+ chunk->ctx = ctx;
+ si_saved_cs_reference(&chunk->cs, scs);
+ chunk->dump_bo_list = dump_bo_list;
+
+ chunk->gfx_begin = scs->gfx_last_dw;
+ chunk->gfx_end = gfx_cur;
+ scs->gfx_last_dw = gfx_cur;
+
+ u_log_chunk(log, &si_log_chunk_type_cs, chunk);
+}
- if (sctx->init_config_gs_rings)
- ac_parse_ib(f, sctx->init_config_gs_rings->pm4,
- sctx->init_config_gs_rings->ndw,
- -1, "IB2: Init GS rings", sctx->b.chip_class,
- NULL, NULL);
+void si_auto_log_cs(void *data, struct u_log_context *log)
+{
+ struct si_context *ctx = (struct si_context *)data;
+ si_log_cs(ctx, log, false);
+}
+
+void si_log_hw_flush(struct si_context *sctx)
+{
+ if (!sctx->log)
+ return;
- ac_parse_ib(f, sctx->last_gfx.ib, sctx->last_gfx.num_dw,
- last_trace_id, "IB", sctx->b.chip_class,
- NULL, NULL);
+ si_log_cs(sctx, sctx->log, true);
}
static const char *priority_to_string(enum radeon_bo_priority priority)
ITEM(IB2),
ITEM(DRAW_INDIRECT),
ITEM(INDEX_BUFFER),
- ITEM(VCE),
- ITEM(UVD),
- ITEM(SDMA_BUFFER),
- ITEM(SDMA_TEXTURE),
ITEM(CP_DMA),
ITEM(CONST_BUFFER),
ITEM(DESCRIPTORS),
ITEM(DEPTH_BUFFER),
ITEM(COLOR_BUFFER_MSAA),
ITEM(DEPTH_BUFFER_MSAA),
- ITEM(CMASK),
- ITEM(DCC),
- ITEM(HTILE),
+ ITEM(SEPARATE_META),
ITEM(SHADER_BINARY),
ITEM(SHADER_RINGS),
ITEM(SCRATCH_BUFFER),
for (i = 0; i < saved->bo_count; i++) {
/* Note: Buffer sizes are expected to be aligned to 4k by the winsys. */
- const unsigned page_size = sctx->b.screen->info.gart_page_size;
+ const unsigned page_size = sctx->screen->info.gart_page_size;
uint64_t va = saved->bo_list[i].vm_address;
uint64_t size = saved->bo_list[i].bo_size;
bool hit = false;
size / page_size, va / page_size, (va + size) / page_size);
/* Print the usage. */
- for (j = 0; j < 64; j++) {
- if (!(saved->bo_list[i].priority_usage & (1ull << j)))
+ for (j = 0; j < 32; j++) {
+ if (!(saved->bo_list[i].priority_usage & (1u << j)))
continue;
fprintf(f, "%s%s", !hit ? "" : ", ", priority_to_string(j));
" Other buffers can still be allocated there.\n\n");
}
-static void si_dump_framebuffer(struct si_context *sctx, FILE *f)
+static void si_dump_framebuffer(struct si_context *sctx, struct u_log_context *log)
{
struct pipe_framebuffer_state *state = &sctx->framebuffer.state;
- struct r600_texture *rtex;
+ struct si_texture *tex;
int i;
for (i = 0; i < state->nr_cbufs; i++) {
if (!state->cbufs[i])
continue;
- rtex = (struct r600_texture*)state->cbufs[i]->texture;
- fprintf(f, COLOR_YELLOW "Color buffer %i:" COLOR_RESET "\n", i);
- r600_print_texture_info(sctx->b.screen, rtex, f);
- fprintf(f, "\n");
+ tex = (struct si_texture*)state->cbufs[i]->texture;
+ u_log_printf(log, COLOR_YELLOW "Color buffer %i:" COLOR_RESET "\n", i);
+ si_print_texture_info(sctx->screen, tex, log);
+ u_log_printf(log, "\n");
}
if (state->zsbuf) {
- rtex = (struct r600_texture*)state->zsbuf->texture;
- fprintf(f, COLOR_YELLOW "Depth-stencil buffer:" COLOR_RESET "\n");
- r600_print_texture_info(sctx->b.screen, rtex, f);
- fprintf(f, "\n");
+ tex = (struct si_texture*)state->zsbuf->texture;
+ u_log_printf(log, COLOR_YELLOW "Depth-stencil buffer:" COLOR_RESET "\n");
+ si_print_texture_info(sctx->screen, tex, log);
+ u_log_printf(log, "\n");
}
}
typedef unsigned (*slot_remap_func)(unsigned);
-static void si_dump_descriptor_list(struct si_descriptors *desc,
- const char *shader_name,
- const char *elem_name,
- unsigned element_dw_size,
- unsigned num_elements,
- slot_remap_func slot_remap,
- FILE *f)
+struct si_log_chunk_desc_list {
+ /** Pointer to memory map of buffer where the list is uploader */
+ uint32_t *gpu_list;
+ /** Reference of buffer where the list is uploaded, so that gpu_list
+ * is kept live. */
+ struct r600_resource *buf;
+
+ const char *shader_name;
+ const char *elem_name;
+ slot_remap_func slot_remap;
+ enum chip_class chip_class;
+ unsigned element_dw_size;
+ unsigned num_elements;
+
+ uint32_t list[0];
+};
+
+static void
+si_log_chunk_desc_list_destroy(void *data)
+{
+ struct si_log_chunk_desc_list *chunk = data;
+ r600_resource_reference(&chunk->buf, NULL);
+ FREE(chunk);
+}
+
+static void
+si_log_chunk_desc_list_print(void *data, FILE *f)
{
- unsigned i, j;
+ struct si_log_chunk_desc_list *chunk = data;
- for (i = 0; i < num_elements; i++) {
- unsigned dw_offset = slot_remap(i) * element_dw_size;
- uint32_t *gpu_ptr = desc->gpu_list ? desc->gpu_list : desc->list;
- const char *list_note = desc->gpu_list ? "GPU list" : "CPU list";
- uint32_t *cpu_list = desc->list + dw_offset;
- uint32_t *gpu_list = gpu_ptr + dw_offset;
+ for (unsigned i = 0; i < chunk->num_elements; i++) {
+ unsigned cpu_dw_offset = i * chunk->element_dw_size;
+ unsigned gpu_dw_offset = chunk->slot_remap(i) * chunk->element_dw_size;
+ const char *list_note = chunk->gpu_list ? "GPU list" : "CPU list";
+ uint32_t *cpu_list = chunk->list + cpu_dw_offset;
+ uint32_t *gpu_list = chunk->gpu_list ? chunk->gpu_list + gpu_dw_offset : cpu_list;
fprintf(f, COLOR_GREEN "%s%s slot %u (%s):" COLOR_RESET "\n",
- shader_name, elem_name, i, list_note);
+ chunk->shader_name, chunk->elem_name, i, list_note);
- switch (element_dw_size) {
+ switch (chunk->element_dw_size) {
case 4:
- for (j = 0; j < 4; j++)
- ac_dump_reg(f, R_008F00_SQ_BUF_RSRC_WORD0 + j*4,
+ for (unsigned j = 0; j < 4; j++)
+ ac_dump_reg(f, chunk->chip_class,
+ R_008F00_SQ_BUF_RSRC_WORD0 + j*4,
gpu_list[j], 0xffffffff);
break;
case 8:
- for (j = 0; j < 8; j++)
- ac_dump_reg(f, R_008F10_SQ_IMG_RSRC_WORD0 + j*4,
+ for (unsigned j = 0; j < 8; j++)
+ ac_dump_reg(f, chunk->chip_class,
+ R_008F10_SQ_IMG_RSRC_WORD0 + j*4,
gpu_list[j], 0xffffffff);
fprintf(f, COLOR_CYAN " Buffer:" COLOR_RESET "\n");
- for (j = 0; j < 4; j++)
- ac_dump_reg(f, R_008F00_SQ_BUF_RSRC_WORD0 + j*4,
+ for (unsigned j = 0; j < 4; j++)
+ ac_dump_reg(f, chunk->chip_class,
+ R_008F00_SQ_BUF_RSRC_WORD0 + j*4,
gpu_list[4+j], 0xffffffff);
break;
case 16:
- for (j = 0; j < 8; j++)
- ac_dump_reg(f, R_008F10_SQ_IMG_RSRC_WORD0 + j*4,
+ for (unsigned j = 0; j < 8; j++)
+ ac_dump_reg(f, chunk->chip_class,
+ R_008F10_SQ_IMG_RSRC_WORD0 + j*4,
gpu_list[j], 0xffffffff);
fprintf(f, COLOR_CYAN " Buffer:" COLOR_RESET "\n");
- for (j = 0; j < 4; j++)
- ac_dump_reg(f, R_008F00_SQ_BUF_RSRC_WORD0 + j*4,
+ for (unsigned j = 0; j < 4; j++)
+ ac_dump_reg(f, chunk->chip_class,
+ R_008F00_SQ_BUF_RSRC_WORD0 + j*4,
gpu_list[4+j], 0xffffffff);
fprintf(f, COLOR_CYAN " FMASK:" COLOR_RESET "\n");
- for (j = 0; j < 8; j++)
- ac_dump_reg(f, R_008F10_SQ_IMG_RSRC_WORD0 + j*4,
+ for (unsigned j = 0; j < 8; j++)
+ ac_dump_reg(f, chunk->chip_class,
+ R_008F10_SQ_IMG_RSRC_WORD0 + j*4,
gpu_list[8+j], 0xffffffff);
fprintf(f, COLOR_CYAN " Sampler state:" COLOR_RESET "\n");
- for (j = 0; j < 4; j++)
- ac_dump_reg(f, R_008F30_SQ_IMG_SAMP_WORD0 + j*4,
+ for (unsigned j = 0; j < 4; j++)
+ ac_dump_reg(f, chunk->chip_class,
+ R_008F30_SQ_IMG_SAMP_WORD0 + j*4,
gpu_list[12+j], 0xffffffff);
break;
}
- if (memcmp(gpu_list, cpu_list, desc->element_dw_size * 4) != 0) {
+ if (memcmp(gpu_list, cpu_list, chunk->element_dw_size * 4) != 0) {
fprintf(f, COLOR_RED "!!!!! This slot was corrupted in GPU memory !!!!!"
COLOR_RESET "\n");
}
fprintf(f, "\n");
}
+
+}
+
+static const struct u_log_chunk_type si_log_chunk_type_descriptor_list = {
+ .destroy = si_log_chunk_desc_list_destroy,
+ .print = si_log_chunk_desc_list_print,
+};
+
+static void si_dump_descriptor_list(struct si_screen *screen,
+ struct si_descriptors *desc,
+ const char *shader_name,
+ const char *elem_name,
+ unsigned element_dw_size,
+ unsigned num_elements,
+ slot_remap_func slot_remap,
+ struct u_log_context *log)
+{
+ if (!desc->list)
+ return;
+
+ /* In some cases, the caller doesn't know how many elements are really
+ * uploaded. Reduce num_elements to fit in the range of active slots. */
+ unsigned active_range_dw_begin =
+ desc->first_active_slot * desc->element_dw_size;
+ unsigned active_range_dw_end =
+ active_range_dw_begin + desc->num_active_slots * desc->element_dw_size;
+
+ while (num_elements > 0) {
+ int i = slot_remap(num_elements - 1);
+ unsigned dw_begin = i * element_dw_size;
+ unsigned dw_end = dw_begin + element_dw_size;
+
+ if (dw_begin >= active_range_dw_begin && dw_end <= active_range_dw_end)
+ break;
+
+ num_elements--;
+ }
+
+ struct si_log_chunk_desc_list *chunk =
+ CALLOC_VARIANT_LENGTH_STRUCT(si_log_chunk_desc_list,
+ 4 * element_dw_size * num_elements);
+ chunk->shader_name = shader_name;
+ chunk->elem_name = elem_name;
+ chunk->element_dw_size = element_dw_size;
+ chunk->num_elements = num_elements;
+ chunk->slot_remap = slot_remap;
+ chunk->chip_class = screen->info.chip_class;
+
+ r600_resource_reference(&chunk->buf, desc->buffer);
+ chunk->gpu_list = desc->gpu_list;
+
+ for (unsigned i = 0; i < num_elements; ++i) {
+ memcpy(&chunk->list[i * element_dw_size],
+ &desc->list[slot_remap(i) * element_dw_size],
+ 4 * element_dw_size);
+ }
+
+ u_log_chunk(log, &si_log_chunk_type_descriptor_list, chunk);
}
static unsigned si_identity(unsigned slot)
static void si_dump_descriptors(struct si_context *sctx,
enum pipe_shader_type processor,
- const struct tgsi_shader_info *info, FILE *f)
+ const struct tgsi_shader_info *info,
+ struct u_log_context *log)
{
struct si_descriptors *descs =
&sctx->descriptors[SI_DESCS_FIRST_SHADER +
u_bit_consecutive(0, SI_NUM_SHADER_BUFFERS);
enabled_shaderbuf = util_bitreverse(enabled_shaderbuf) >>
(32 - SI_NUM_SHADER_BUFFERS);
- enabled_samplers = sctx->samplers[processor].views.enabled_mask;
+ enabled_samplers = sctx->samplers[processor].enabled_mask;
enabled_images = sctx->images[processor].enabled_mask;
}
- if (processor == PIPE_SHADER_VERTEX) {
+ if (processor == PIPE_SHADER_VERTEX &&
+ sctx->vb_descriptors_buffer &&
+ sctx->vb_descriptors_gpu_list &&
+ sctx->vertex_elements) {
assert(info); /* only CS may not have an info struct */
+ struct si_descriptors desc = {};
+
+ desc.buffer = sctx->vb_descriptors_buffer;
+ desc.list = sctx->vb_descriptors_gpu_list;
+ desc.gpu_list = sctx->vb_descriptors_gpu_list;
+ desc.element_dw_size = 4;
+ desc.num_active_slots = sctx->vertex_elements->desc_list_byte_size / 16;
- si_dump_descriptor_list(&sctx->vertex_buffers, name,
+ si_dump_descriptor_list(sctx->screen, &desc, name,
" - Vertex buffer", 4, info->num_inputs,
- si_identity, f);
+ si_identity, log);
}
- si_dump_descriptor_list(&descs[SI_SHADER_DESCS_CONST_AND_SHADER_BUFFERS],
+ si_dump_descriptor_list(sctx->screen,
+ &descs[SI_SHADER_DESCS_CONST_AND_SHADER_BUFFERS],
name, " - Constant buffer", 4,
util_last_bit(enabled_constbuf),
- si_get_constbuf_slot, f);
- si_dump_descriptor_list(&descs[SI_SHADER_DESCS_CONST_AND_SHADER_BUFFERS],
+ si_get_constbuf_slot, log);
+ si_dump_descriptor_list(sctx->screen,
+ &descs[SI_SHADER_DESCS_CONST_AND_SHADER_BUFFERS],
name, " - Shader buffer", 4,
util_last_bit(enabled_shaderbuf),
- si_get_shaderbuf_slot, f);
- si_dump_descriptor_list(&descs[SI_SHADER_DESCS_SAMPLERS_AND_IMAGES],
+ si_get_shaderbuf_slot, log);
+ si_dump_descriptor_list(sctx->screen,
+ &descs[SI_SHADER_DESCS_SAMPLERS_AND_IMAGES],
name, " - Sampler", 16,
util_last_bit(enabled_samplers),
- si_get_sampler_slot, f);
- si_dump_descriptor_list(&descs[SI_SHADER_DESCS_SAMPLERS_AND_IMAGES],
+ si_get_sampler_slot, log);
+ si_dump_descriptor_list(sctx->screen,
+ &descs[SI_SHADER_DESCS_SAMPLERS_AND_IMAGES],
name, " - Image", 8,
util_last_bit(enabled_images),
- si_get_image_slot, f);
+ si_get_image_slot, log);
}
static void si_dump_gfx_descriptors(struct si_context *sctx,
const struct si_shader_ctx_state *state,
- FILE *f)
+ struct u_log_context *log)
{
if (!state->cso || !state->current)
return;
- si_dump_descriptors(sctx, state->cso->type, &state->cso->info, f);
+ si_dump_descriptors(sctx, state->cso->type, &state->cso->info, log);
}
-static void si_dump_compute_descriptors(struct si_context *sctx, FILE *f)
+static void si_dump_compute_descriptors(struct si_context *sctx,
+ struct u_log_context *log)
{
- if (!sctx->cs_shader_state.program ||
- sctx->cs_shader_state.program != sctx->cs_shader_state.emitted_program)
+ if (!sctx->cs_shader_state.program)
return;
- si_dump_descriptors(sctx, PIPE_SHADER_COMPUTE, NULL, f);
+ si_dump_descriptors(sctx, PIPE_SHADER_COMPUTE, NULL, log);
}
struct si_shader_inst {
- char text[160]; /* one disasm line */
- unsigned offset; /* instruction offset */
+ const char *text; /* start of disassembly for this instruction */
+ unsigned textlen;
unsigned size; /* instruction size = 4 or 8 */
+ uint64_t addr; /* instruction address */
};
-/* Split a disassembly string into lines and add them to the array pointed
- * to by "instructions". */
+/**
+ * Split a disassembly string into instructions and add them to the array
+ * pointed to by \p instructions.
+ *
+ * Labels are considered to be part of the following instruction.
+ */
static void si_add_split_disasm(const char *disasm,
- uint64_t start_addr,
+ uint64_t *addr,
unsigned *num,
struct si_shader_inst *instructions)
{
- struct si_shader_inst *last_inst = *num ? &instructions[*num - 1] : NULL;
- char *next;
+ const char *semicolon;
- while ((next = strchr(disasm, '\n'))) {
- struct si_shader_inst *inst = &instructions[*num];
- unsigned len = next - disasm;
+ while ((semicolon = strchr(disasm, ';'))) {
+ struct si_shader_inst *inst = &instructions[(*num)++];
+ const char *end = util_strchrnul(semicolon, '\n');
- assert(len < ARRAY_SIZE(inst->text));
- memcpy(inst->text, disasm, len);
- inst->text[len] = 0;
- inst->offset = last_inst ? last_inst->offset + last_inst->size : 0;
+ inst->text = disasm;
+ inst->textlen = end - disasm;
- const char *semicolon = strchr(disasm, ';');
- assert(semicolon);
+ inst->addr = *addr;
/* More than 16 chars after ";" means the instruction is 8 bytes long. */
- inst->size = next - semicolon > 16 ? 8 : 4;
-
- snprintf(inst->text + len, ARRAY_SIZE(inst->text) - len,
- " [PC=0x%"PRIx64", off=%u, size=%u]",
- start_addr + inst->offset, inst->offset, inst->size);
-
- last_inst = inst;
- (*num)++;
- disasm = next + 1;
- }
-}
-
-#define MAX_WAVES_PER_CHIP (64 * 40)
-
-struct si_wave_info {
- unsigned se; /* shader engine */
- unsigned sh; /* shader array */
- unsigned cu; /* compute unit */
- unsigned simd;
- unsigned wave;
- uint32_t status;
- uint64_t pc; /* program counter */
- uint32_t inst_dw0;
- uint32_t inst_dw1;
- uint64_t exec;
- bool matched; /* whether the wave is used by a currently-bound shader */
-};
-
-static int compare_wave(const void *p1, const void *p2)
-{
- struct si_wave_info *w1 = (struct si_wave_info *)p1;
- struct si_wave_info *w2 = (struct si_wave_info *)p2;
-
- /* Sort waves according to PC and then SE, SH, CU, etc. */
- if (w1->pc < w2->pc)
- return -1;
- if (w1->pc > w2->pc)
- return 1;
- if (w1->se < w2->se)
- return -1;
- if (w1->se > w2->se)
- return 1;
- if (w1->sh < w2->sh)
- return -1;
- if (w1->sh > w2->sh)
- return 1;
- if (w1->cu < w2->cu)
- return -1;
- if (w1->cu > w2->cu)
- return 1;
- if (w1->simd < w2->simd)
- return -1;
- if (w1->simd > w2->simd)
- return 1;
- if (w1->wave < w2->wave)
- return -1;
- if (w1->wave > w2->wave)
- return 1;
-
- return 0;
-}
-
-/* Return wave information. "waves" should be a large enough array. */
-static unsigned si_get_wave_info(struct si_wave_info waves[MAX_WAVES_PER_CHIP])
-{
- char line[2000];
- unsigned num_waves = 0;
-
- FILE *p = popen("umr -wa", "r");
- if (!p)
- return 0;
+ inst->size = end - semicolon > 16 ? 8 : 4;
+ *addr += inst->size;
- if (!fgets(line, sizeof(line), p) ||
- strncmp(line, "SE", 2) != 0) {
- pclose(p);
- return 0;
- }
-
- while (fgets(line, sizeof(line), p)) {
- struct si_wave_info *w;
- uint32_t pc_hi, pc_lo, exec_hi, exec_lo;
-
- assert(num_waves < MAX_WAVES_PER_CHIP);
- w = &waves[num_waves];
-
- if (sscanf(line, "%u %u %u %u %u %x %x %x %x %x %x %x",
- &w->se, &w->sh, &w->cu, &w->simd, &w->wave,
- &w->status, &pc_hi, &pc_lo, &w->inst_dw0,
- &w->inst_dw1, &exec_hi, &exec_lo) == 12) {
- w->pc = ((uint64_t)pc_hi << 32) | pc_lo;
- w->exec = ((uint64_t)exec_hi << 32) | exec_lo;
- w->matched = false;
- num_waves++;
- }
+ if (!(*end))
+ break;
+ disasm = end + 1;
}
-
- qsort(waves, num_waves, sizeof(struct si_wave_info), compare_wave);
-
- pclose(p);
- return num_waves;
}
/* If the shader is being executed, print its asm instructions, and annotate
* execute them. This is most useful during a GPU hang.
*/
static void si_print_annotated_shader(struct si_shader *shader,
- struct si_wave_info *waves,
+ struct ac_wave_info *waves,
unsigned num_waves,
FILE *f)
{
* Buffer size / 4 is the upper bound of the instruction count.
*/
unsigned num_inst = 0;
+ uint64_t inst_addr = start_addr;
struct si_shader_inst *instructions =
calloc(shader->bo->b.b.width0 / 4, sizeof(struct si_shader_inst));
if (shader->prolog) {
si_add_split_disasm(shader->prolog->binary.disasm_string,
- start_addr, &num_inst, instructions);
+ &inst_addr, &num_inst, instructions);
}
if (shader->previous_stage) {
si_add_split_disasm(shader->previous_stage->binary.disasm_string,
- start_addr, &num_inst, instructions);
+ &inst_addr, &num_inst, instructions);
}
if (shader->prolog2) {
si_add_split_disasm(shader->prolog2->binary.disasm_string,
- start_addr, &num_inst, instructions);
+ &inst_addr, &num_inst, instructions);
}
si_add_split_disasm(shader->binary.disasm_string,
- start_addr, &num_inst, instructions);
+ &inst_addr, &num_inst, instructions);
if (shader->epilog) {
si_add_split_disasm(shader->epilog->binary.disasm_string,
- start_addr, &num_inst, instructions);
+ &inst_addr, &num_inst, instructions);
}
fprintf(f, COLOR_YELLOW "%s - annotated disassembly:" COLOR_RESET "\n",
for (i = 0; i < num_inst; i++) {
struct si_shader_inst *inst = &instructions[i];
- fprintf(f, "%s\n", inst->text);
+ fprintf(f, "%.*s [PC=0x%"PRIx64", size=%u]\n",
+ inst->textlen, inst->text, inst->addr, inst->size);
/* Print which waves execute the instruction right now. */
- while (num_waves && start_addr + inst->offset == waves->pc) {
+ while (num_waves && inst->addr == waves->pc) {
fprintf(f,
" " COLOR_GREEN "^ SE%u SH%u CU%u "
"SIMD%u WAVE%u EXEC=%016"PRIx64 " ",
static void si_dump_annotated_shaders(struct si_context *sctx, FILE *f)
{
- struct si_wave_info waves[MAX_WAVES_PER_CHIP];
- unsigned num_waves = si_get_wave_info(waves);
+ struct ac_wave_info waves[AC_MAX_WAVES_PER_CHIP];
+ unsigned num_waves = ac_get_wave_info(waves);
fprintf(f, COLOR_CYAN "The number of active waves = %u" COLOR_RESET
"\n\n", num_waves);
{
struct si_context *sctx = (struct si_context*)ctx;
- if (flags & PIPE_DUMP_DEVICE_STATUS_REGISTERS)
- si_dump_debug_registers(sctx, f);
+ if (sctx->log)
+ u_log_flush(sctx->log);
- if (flags & PIPE_DUMP_CURRENT_STATES)
- si_dump_framebuffer(sctx, f);
-
- if (flags & PIPE_DUMP_CURRENT_SHADERS) {
- si_dump_gfx_shader(sctx->screen, &sctx->vs_shader, f);
- si_dump_gfx_shader(sctx->screen, &sctx->tcs_shader, f);
- si_dump_gfx_shader(sctx->screen, &sctx->tes_shader, f);
- si_dump_gfx_shader(sctx->screen, &sctx->gs_shader, f);
- si_dump_gfx_shader(sctx->screen, &sctx->ps_shader, f);
- si_dump_compute_shader(sctx->screen, &sctx->cs_shader_state, f);
-
- if (flags & PIPE_DUMP_DEVICE_STATUS_REGISTERS) {
- si_dump_annotated_shaders(sctx, f);
- si_dump_command("Active waves (raw data)", "umr -wa | column -t", f);
- si_dump_command("Wave information", "umr -O bits -wa", f);
- }
+ if (flags & PIPE_DUMP_DEVICE_STATUS_REGISTERS) {
+ si_dump_debug_registers(sctx, f);
- si_dump_descriptor_list(&sctx->descriptors[SI_DESCS_RW_BUFFERS],
- "", "RW buffers", 4, SI_NUM_RW_BUFFERS,
- si_identity, f);
- si_dump_gfx_descriptors(sctx, &sctx->vs_shader, f);
- si_dump_gfx_descriptors(sctx, &sctx->tcs_shader, f);
- si_dump_gfx_descriptors(sctx, &sctx->tes_shader, f);
- si_dump_gfx_descriptors(sctx, &sctx->gs_shader, f);
- si_dump_gfx_descriptors(sctx, &sctx->ps_shader, f);
- si_dump_compute_descriptors(sctx, f);
+ si_dump_annotated_shaders(sctx, f);
+ si_dump_command("Active waves (raw data)", "umr -O halt_waves -wa | column -t", f);
+ si_dump_command("Wave information", "umr -O halt_waves,bits -wa", f);
}
+}
- if (flags & PIPE_DUMP_LAST_COMMAND_BUFFER) {
- si_dump_bo_list(sctx, &sctx->last_gfx, f);
- si_dump_last_ib(sctx, f);
+void si_log_draw_state(struct si_context *sctx, struct u_log_context *log)
+{
+ struct si_shader_ctx_state *tcs_shader;
- fprintf(f, "Done.\n");
+ if (!log)
+ return;
- /* dump only once */
- radeon_clear_saved_cs(&sctx->last_gfx);
- r600_resource_reference(&sctx->last_trace_buf, NULL);
- }
+ tcs_shader = &sctx->tcs_shader;
+ if (sctx->tes_shader.cso && !sctx->tcs_shader.cso)
+ tcs_shader = &sctx->fixed_func_tcs_shader;
+
+ si_dump_framebuffer(sctx, log);
+
+ si_dump_gfx_shader(sctx, &sctx->vs_shader, log);
+ si_dump_gfx_shader(sctx, tcs_shader, log);
+ si_dump_gfx_shader(sctx, &sctx->tes_shader, log);
+ si_dump_gfx_shader(sctx, &sctx->gs_shader, log);
+ si_dump_gfx_shader(sctx, &sctx->ps_shader, log);
+
+ si_dump_descriptor_list(sctx->screen,
+ &sctx->descriptors[SI_DESCS_RW_BUFFERS],
+ "", "RW buffers", 4, SI_NUM_RW_BUFFERS,
+ si_identity, log);
+ si_dump_gfx_descriptors(sctx, &sctx->vs_shader, log);
+ si_dump_gfx_descriptors(sctx, tcs_shader, log);
+ si_dump_gfx_descriptors(sctx, &sctx->tes_shader, log);
+ si_dump_gfx_descriptors(sctx, &sctx->gs_shader, log);
+ si_dump_gfx_descriptors(sctx, &sctx->ps_shader, log);
+}
+
+void si_log_compute_state(struct si_context *sctx, struct u_log_context *log)
+{
+ if (!log)
+ return;
+
+ si_dump_compute_shader(sctx, log);
+ si_dump_compute_descriptors(sctx, log);
}
static void si_dump_dma(struct si_context *sctx,
fprintf(f, "SDMA Dump Done.\n");
}
-static bool si_vm_fault_occured(struct si_context *sctx, uint64_t *out_addr)
-{
- char line[2000];
- unsigned sec, usec;
- int progress = 0;
- uint64_t timestamp = 0;
- bool fault = false;
-
- FILE *p = popen("dmesg", "r");
- if (!p)
- return false;
-
- while (fgets(line, sizeof(line), p)) {
- char *msg, len;
-
- if (!line[0] || line[0] == '\n')
- continue;
-
- /* Get the timestamp. */
- if (sscanf(line, "[%u.%u]", &sec, &usec) != 2) {
- static bool hit = false;
- if (!hit) {
- fprintf(stderr, "%s: failed to parse line '%s'\n",
- __func__, line);
- hit = true;
- }
- continue;
- }
- timestamp = sec * 1000000ull + usec;
-
- /* If just updating the timestamp. */
- if (!out_addr)
- continue;
-
- /* Process messages only if the timestamp is newer. */
- if (timestamp <= sctx->dmesg_timestamp)
- continue;
-
- /* Only process the first VM fault. */
- if (fault)
- continue;
-
- /* Remove trailing \n */
- len = strlen(line);
- if (len && line[len-1] == '\n')
- line[len-1] = 0;
-
- /* Get the message part. */
- msg = strchr(line, ']');
- if (!msg) {
- assert(0);
- continue;
- }
- msg++;
-
- const char *header_line, *addr_line_prefix, *addr_line_format;
-
- if (sctx->b.chip_class >= GFX9) {
- /* Match this:
- * ..: [gfxhub] VMC page fault (src_id:0 ring:158 vm_id:2 pas_id:0)
- * ..: at page 0x0000000219f8f000 from 27
- * ..: VM_L2_PROTECTION_FAULT_STATUS:0x0020113C
- */
- header_line = "VMC page fault";
- addr_line_prefix = " at page";
- addr_line_format = "%"PRIx64;
- } else {
- header_line = "GPU fault detected:";
- addr_line_prefix = "VM_CONTEXT1_PROTECTION_FAULT_ADDR";
- addr_line_format = "%"PRIX64;
- }
-
- switch (progress) {
- case 0:
- if (strstr(msg, header_line))
- progress = 1;
- break;
- case 1:
- msg = strstr(msg, addr_line_prefix);
- if (msg) {
- msg = strstr(msg, "0x");
- if (msg) {
- msg += 2;
- if (sscanf(msg, addr_line_format, out_addr) == 1)
- fault = true;
- }
- }
- progress = 0;
- break;
- default:
- progress = 0;
- }
- }
- pclose(p);
-
- if (timestamp > sctx->dmesg_timestamp)
- sctx->dmesg_timestamp = timestamp;
- return fault;
-}
-
-void si_check_vm_faults(struct r600_common_context *ctx,
+void si_check_vm_faults(struct si_context *sctx,
struct radeon_saved_cs *saved, enum ring_type ring)
{
- struct si_context *sctx = (struct si_context *)ctx;
- struct pipe_screen *screen = sctx->b.b.screen;
+ struct pipe_screen *screen = sctx->b.screen;
FILE *f;
uint64_t addr;
char cmd_line[4096];
- if (!si_vm_fault_occured(sctx, &addr))
+ if (!ac_vm_fault_occured(sctx->chip_class,
+ &sctx->dmesg_timestamp, &addr))
return;
f = dd_get_debug_file(false);
sctx->apitrace_call_number);
switch (ring) {
- case RING_GFX:
- si_dump_debug_state(&sctx->b.b, f,
- PIPE_DUMP_CURRENT_STATES |
- PIPE_DUMP_CURRENT_SHADERS |
- PIPE_DUMP_LAST_COMMAND_BUFFER);
- break;
+ case RING_GFX: {
+ struct u_log_context log;
+ u_log_context_init(&log);
+ si_log_draw_state(sctx, &log);
+ si_log_compute_state(sctx, &log);
+ si_log_cs(sctx, &log, true);
+
+ u_log_new_page_print(&log, f);
+ u_log_context_destroy(&log);
+ break;
+ }
case RING_DMA:
si_dump_dma(sctx, saved, f);
break;
void si_init_debug_functions(struct si_context *sctx)
{
- sctx->b.b.dump_debug_state = si_dump_debug_state;
- sctx->b.check_vm_faults = si_check_vm_faults;
+ sctx->b.dump_debug_state = si_dump_debug_state;
/* Set the initial dmesg timestamp for this context, so that
* only new messages will be checked for VM faults.
*/
- if (sctx->screen->b.debug_flags & DBG_CHECK_VM)
- si_vm_fault_occured(sctx, NULL);
+ if (sctx->screen->debug_flags & DBG(CHECK_VM))
+ ac_vm_fault_occured(sctx->chip_class,
+ &sctx->dmesg_timestamp, NULL);
}
projects / mesa.git / blobdiff