case R600_QUERY_NUM_SHADERS_CREATED:
query->begin_result = p_atomic_read(&rctx->screen->num_shaders_created);
break;
+ case R600_QUERY_GPIN_ASIC_ID:
+ case R600_QUERY_GPIN_NUM_SIMD:
+ case R600_QUERY_GPIN_NUM_RB:
+ case R600_QUERY_GPIN_NUM_SPI:
+ case R600_QUERY_GPIN_NUM_SE:
+ break;
default:
unreachable("r600_query_sw_begin: bad query type");
}
rctx->b.flush(&rctx->b, &query->fence, 0);
break;
case R600_QUERY_DRAW_CALLS:
- query->begin_result = rctx->num_draw_calls;
+ query->end_result = rctx->num_draw_calls;
break;
case R600_QUERY_REQUESTED_VRAM:
case R600_QUERY_REQUESTED_GTT:
query->begin_result = 0;
break;
case R600_QUERY_NUM_COMPILATIONS:
- query->begin_result = p_atomic_read(&rctx->screen->num_compilations);
+ query->end_result = p_atomic_read(&rctx->screen->num_compilations);
break;
case R600_QUERY_NUM_SHADERS_CREATED:
- query->begin_result = p_atomic_read(&rctx->screen->num_shaders_created);
+ query->end_result = p_atomic_read(&rctx->screen->num_shaders_created);
+ break;
+ case R600_QUERY_GPIN_ASIC_ID:
+ case R600_QUERY_GPIN_NUM_SIMD:
+ case R600_QUERY_GPIN_NUM_RB:
+ case R600_QUERY_GPIN_NUM_SPI:
+ case R600_QUERY_GPIN_NUM_SE:
break;
default:
unreachable("r600_query_sw_end: bad query type");
case PIPE_QUERY_TIMESTAMP_DISJOINT:
/* Convert from cycles per millisecond to cycles per second (Hz). */
result->timestamp_disjoint.frequency =
- (uint64_t)rctx->screen->info.r600_clock_crystal_freq * 1000;
+ (uint64_t)rctx->screen->info.clock_crystal_freq * 1000;
result->timestamp_disjoint.disjoint = FALSE;
return TRUE;
case PIPE_QUERY_GPU_FINISHED: {
wait ? PIPE_TIMEOUT_INFINITE : 0);
return result->b;
}
+
+ case R600_QUERY_GPIN_ASIC_ID:
+ result->u32 = 0;
+ return TRUE;
+ case R600_QUERY_GPIN_NUM_SIMD:
+ result->u32 = rctx->screen->info.num_good_compute_units;
+ return TRUE;
+ case R600_QUERY_GPIN_NUM_RB:
+ result->u32 = rctx->screen->info.num_render_backends;
+ return TRUE;
+ case R600_QUERY_GPIN_NUM_SPI:
+ result->u32 = 1; /* all supported chips have one SPI per SE */
+ return TRUE;
+ case R600_QUERY_GPIN_NUM_SE:
+ result->u32 = rctx->screen->info.max_se;
+ return TRUE;
}
result->u64 = query->end_result - query->begin_result;
struct r600_query_sw *query;
query = CALLOC_STRUCT(r600_query_sw);
- if (query == NULL)
+ if (!query)
return NULL;
query->b.type = query_type;
FREE(rquery);
}
-static bool r600_is_timer_query(unsigned type)
+static struct r600_resource *r600_new_query_buffer(struct r600_common_context *ctx,
+ struct r600_query_hw *query)
{
- return type == PIPE_QUERY_TIME_ELAPSED ||
- type == PIPE_QUERY_TIMESTAMP;
-}
-
-static bool r600_query_needs_begin(unsigned type)
-{
- return type != PIPE_QUERY_TIMESTAMP;
-}
-
-static struct r600_resource *r600_new_query_buffer(struct r600_common_context *ctx, unsigned type)
-{
- unsigned j, i, num_results, buf_size = 4096;
- uint32_t *results;
+ unsigned buf_size = MAX2(query->result_size, 4096);
/* Queries are normally read by the CPU after
* being written by the gpu, hence staging is probably a good
pipe_buffer_create(ctx->b.screen, PIPE_BIND_CUSTOM,
PIPE_USAGE_STAGING, buf_size);
- switch (type) {
- case PIPE_QUERY_OCCLUSION_COUNTER:
- case PIPE_QUERY_OCCLUSION_PREDICATE:
- results = r600_buffer_map_sync_with_rings(ctx, buf, PIPE_TRANSFER_WRITE);
- memset(results, 0, buf_size);
+ if (query->flags & R600_QUERY_HW_FLAG_PREDICATE)
+ query->ops->prepare_buffer(ctx, query, buf);
+
+ return buf;
+}
+
+static void r600_query_hw_prepare_buffer(struct r600_common_context *ctx,
+ struct r600_query_hw *query,
+ struct r600_resource *buffer)
+{
+ /* Callers ensure that the buffer is currently unused by the GPU. */
+ uint32_t *results = ctx->ws->buffer_map(buffer->buf, NULL,
+ PIPE_TRANSFER_WRITE |
+ PIPE_TRANSFER_UNSYNCHRONIZED);
+
+ memset(results, 0, buffer->b.b.width0);
+
+ if (query->b.type == PIPE_QUERY_OCCLUSION_COUNTER ||
+ query->b.type == PIPE_QUERY_OCCLUSION_PREDICATE) {
+ unsigned num_results;
+ unsigned i, j;
/* Set top bits for unused backends. */
- num_results = buf_size / (16 * ctx->max_db);
+ num_results = buffer->b.b.width0 / (16 * ctx->max_db);
for (j = 0; j < num_results; j++) {
for (i = 0; i < ctx->max_db; i++) {
if (!(ctx->backend_mask & (1<<i))) {
}
results += 4 * ctx->max_db;
}
- break;
- case PIPE_QUERY_TIME_ELAPSED:
- case PIPE_QUERY_TIMESTAMP:
- break;
- case PIPE_QUERY_PRIMITIVES_EMITTED:
- case PIPE_QUERY_PRIMITIVES_GENERATED:
- case PIPE_QUERY_SO_STATISTICS:
- case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
- case PIPE_QUERY_PIPELINE_STATISTICS:
- results = r600_buffer_map_sync_with_rings(ctx, buf, PIPE_TRANSFER_WRITE);
- memset(results, 0, buf_size);
- break;
- default:
- assert(0);
}
- return buf;
}
-static boolean r600_query_hw_begin(struct r600_common_context *, struct r600_query *);
-static void r600_query_hw_end(struct r600_common_context *, struct r600_query *);
-static boolean r600_query_hw_get_result(struct r600_common_context *,
- struct r600_query *, boolean wait,
- union pipe_query_result *result);
-
static struct r600_query_ops query_hw_ops = {
.destroy = r600_query_hw_destroy,
.begin = r600_query_hw_begin,
.get_result = r600_query_hw_get_result,
};
+static void r600_query_hw_do_emit_start(struct r600_common_context *ctx,
+ struct r600_query_hw *query,
+ struct r600_resource *buffer,
+ uint64_t va);
+static void r600_query_hw_do_emit_stop(struct r600_common_context *ctx,
+ struct r600_query_hw *query,
+ struct r600_resource *buffer,
+ uint64_t va);
+static void r600_query_hw_add_result(struct r600_common_context *ctx,
+ struct r600_query_hw *, void *buffer,
+ union pipe_query_result *result);
+static void r600_query_hw_clear_result(struct r600_query_hw *,
+ union pipe_query_result *);
+
+static struct r600_query_hw_ops query_hw_default_hw_ops = {
+ .prepare_buffer = r600_query_hw_prepare_buffer,
+ .emit_start = r600_query_hw_do_emit_start,
+ .emit_stop = r600_query_hw_do_emit_stop,
+ .clear_result = r600_query_hw_clear_result,
+ .add_result = r600_query_hw_add_result,
+};
+
+boolean r600_query_hw_init(struct r600_common_context *rctx,
+ struct r600_query_hw *query)
+{
+ query->buffer.buf = r600_new_query_buffer(rctx, query);
+ if (!query->buffer.buf)
+ return FALSE;
+
+ return TRUE;
+}
+
static struct pipe_query *r600_query_hw_create(struct r600_common_context *rctx,
unsigned query_type,
unsigned index)
query->b.type = query_type;
query->b.ops = &query_hw_ops;
+ query->ops = &query_hw_default_hw_ops;
switch (query_type) {
case PIPE_QUERY_OCCLUSION_COUNTER:
case PIPE_QUERY_OCCLUSION_PREDICATE:
query->result_size = 16 * rctx->max_db;
- query->num_cs_dw = 6;
+ query->num_cs_dw_begin = 6;
+ query->num_cs_dw_end = 6;
+ query->flags |= R600_QUERY_HW_FLAG_PREDICATE;
break;
case PIPE_QUERY_TIME_ELAPSED:
query->result_size = 16;
- query->num_cs_dw = 8;
+ query->num_cs_dw_begin = 8;
+ query->num_cs_dw_end = 8;
+ query->flags = R600_QUERY_HW_FLAG_TIMER;
break;
case PIPE_QUERY_TIMESTAMP:
query->result_size = 8;
- query->num_cs_dw = 8;
+ query->num_cs_dw_end = 8;
+ query->flags = R600_QUERY_HW_FLAG_TIMER |
+ R600_QUERY_HW_FLAG_NO_START;
break;
case PIPE_QUERY_PRIMITIVES_EMITTED:
case PIPE_QUERY_PRIMITIVES_GENERATED:
case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
/* NumPrimitivesWritten, PrimitiveStorageNeeded. */
query->result_size = 32;
- query->num_cs_dw = 6;
+ query->num_cs_dw_begin = 6;
+ query->num_cs_dw_end = 6;
query->stream = index;
+ query->flags |= R600_QUERY_HW_FLAG_PREDICATE;
break;
case PIPE_QUERY_PIPELINE_STATISTICS:
/* 11 values on EG, 8 on R600. */
query->result_size = (rctx->chip_class >= EVERGREEN ? 11 : 8) * 16;
- query->num_cs_dw = 6;
+ query->num_cs_dw_begin = 6;
+ query->num_cs_dw_end = 6;
break;
default:
assert(0);
return NULL;
}
- query->buffer.buf = r600_new_query_buffer(rctx, query_type);
- if (!query->buffer.buf) {
+ if (!r600_query_hw_init(rctx, query)) {
FREE(query);
return NULL;
}
}
}
-static void r600_emit_query_begin(struct r600_common_context *ctx,
- struct r600_query_hw *query)
+static void r600_query_hw_do_emit_start(struct r600_common_context *ctx,
+ struct r600_query_hw *query,
+ struct r600_resource *buffer,
+ uint64_t va)
{
struct radeon_winsys_cs *cs = ctx->gfx.cs;
- uint64_t va;
-
- r600_update_occlusion_query_state(ctx, query->b.type, 1);
- r600_update_prims_generated_query_state(ctx, query->b.type, 1);
- ctx->need_gfx_cs_space(&ctx->b, query->num_cs_dw * 2, TRUE);
-
- /* Get a new query buffer if needed. */
- if (query->buffer.results_end + query->result_size > query->buffer.buf->b.b.width0) {
- struct r600_query_buffer *qbuf = MALLOC_STRUCT(r600_query_buffer);
- *qbuf = query->buffer;
- query->buffer.buf = r600_new_query_buffer(ctx, query->b.type);
- query->buffer.results_end = 0;
- query->buffer.previous = qbuf;
- }
-
- /* emit begin query */
- va = query->buffer.buf->gpu_address + query->buffer.results_end;
switch (query->b.type) {
case PIPE_QUERY_OCCLUSION_COUNTER:
}
r600_emit_reloc(ctx, &ctx->gfx, query->buffer.buf, RADEON_USAGE_WRITE,
RADEON_PRIO_QUERY);
-
- if (r600_is_timer_query(query->b.type))
- ctx->num_cs_dw_timer_queries_suspend += query->num_cs_dw;
- else
- ctx->num_cs_dw_nontimer_queries_suspend += query->num_cs_dw;
}
-static void r600_emit_query_end(struct r600_common_context *ctx,
- struct r600_query_hw *query)
+static void r600_query_hw_emit_start(struct r600_common_context *ctx,
+ struct r600_query_hw *query)
{
- struct radeon_winsys_cs *cs = ctx->gfx.cs;
uint64_t va;
- /* The queries which need begin already called this in begin_query. */
- if (!r600_query_needs_begin(query->b.type)) {
- ctx->need_gfx_cs_space(&ctx->b, query->num_cs_dw, FALSE);
+ r600_update_occlusion_query_state(ctx, query->b.type, 1);
+ r600_update_prims_generated_query_state(ctx, query->b.type, 1);
+
+ ctx->need_gfx_cs_space(&ctx->b, query->num_cs_dw_begin + query->num_cs_dw_end,
+ TRUE);
+
+ /* Get a new query buffer if needed. */
+ if (query->buffer.results_end + query->result_size > query->buffer.buf->b.b.width0) {
+ struct r600_query_buffer *qbuf = MALLOC_STRUCT(r600_query_buffer);
+ *qbuf = query->buffer;
+ query->buffer.buf = r600_new_query_buffer(ctx, query);
+ query->buffer.results_end = 0;
+ query->buffer.previous = qbuf;
}
- va = query->buffer.buf->gpu_address;
+ /* emit begin query */
+ va = query->buffer.buf->gpu_address + query->buffer.results_end;
+
+ query->ops->emit_start(ctx, query, query->buffer.buf, va);
+
+ if (query->flags & R600_QUERY_HW_FLAG_TIMER)
+ ctx->num_cs_dw_timer_queries_suspend += query->num_cs_dw_end;
+ else
+ ctx->num_cs_dw_nontimer_queries_suspend += query->num_cs_dw_end;
+}
+
+static void r600_query_hw_do_emit_stop(struct r600_common_context *ctx,
+ struct r600_query_hw *query,
+ struct r600_resource *buffer,
+ uint64_t va)
+{
+ struct radeon_winsys_cs *cs = ctx->gfx.cs;
- /* emit end query */
switch (query->b.type) {
case PIPE_QUERY_OCCLUSION_COUNTER:
case PIPE_QUERY_OCCLUSION_PREDICATE:
- va += query->buffer.results_end + 8;
+ va += 8;
radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_ZPASS_DONE) | EVENT_INDEX(1));
radeon_emit(cs, va);
case PIPE_QUERY_PRIMITIVES_GENERATED:
case PIPE_QUERY_SO_STATISTICS:
case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
- va += query->buffer.results_end + query->result_size/2;
+ va += query->result_size/2;
radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
radeon_emit(cs, EVENT_TYPE(event_type_for_stream(query)) | EVENT_INDEX(3));
radeon_emit(cs, va);
radeon_emit(cs, (va >> 32) & 0xFFFF);
break;
case PIPE_QUERY_TIME_ELAPSED:
- va += query->buffer.results_end + query->result_size/2;
+ va += query->result_size/2;
/* fall through */
case PIPE_QUERY_TIMESTAMP:
radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0));
radeon_emit(cs, 0);
break;
case PIPE_QUERY_PIPELINE_STATISTICS:
- va += query->buffer.results_end + query->result_size/2;
+ va += query->result_size/2;
radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_SAMPLE_PIPELINESTAT) | EVENT_INDEX(2));
radeon_emit(cs, va);
}
r600_emit_reloc(ctx, &ctx->gfx, query->buffer.buf, RADEON_USAGE_WRITE,
RADEON_PRIO_QUERY);
+}
+
+static void r600_query_hw_emit_stop(struct r600_common_context *ctx,
+ struct r600_query_hw *query)
+{
+ uint64_t va;
+
+ /* The queries which need begin already called this in begin_query. */
+ if (query->flags & R600_QUERY_HW_FLAG_NO_START) {
+ ctx->need_gfx_cs_space(&ctx->b, query->num_cs_dw_end, FALSE);
+ }
+
+ /* emit end query */
+ va = query->buffer.buf->gpu_address + query->buffer.results_end;
+
+ query->ops->emit_stop(ctx, query, query->buffer.buf, va);
query->buffer.results_end += query->result_size;
- if (r600_query_needs_begin(query->b.type)) {
- if (r600_is_timer_query(query->b.type))
- ctx->num_cs_dw_timer_queries_suspend -= query->num_cs_dw;
+ if (!(query->flags & R600_QUERY_HW_FLAG_NO_START)) {
+ if (query->flags & R600_QUERY_HW_FLAG_TIMER)
+ ctx->num_cs_dw_timer_queries_suspend -= query->num_cs_dw_end;
else
- ctx->num_cs_dw_nontimer_queries_suspend -= query->num_cs_dw;
+ ctx->num_cs_dw_nontimer_queries_suspend -= query->num_cs_dw_end;
}
r600_update_occlusion_query_state(ctx, query->b.type, -1);
return rquery->ops->begin(rctx, rquery);
}
-static boolean r600_query_hw_begin(struct r600_common_context *rctx,
- struct r600_query *rquery)
+static void r600_query_hw_reset_buffers(struct r600_common_context *rctx,
+ struct r600_query_hw *query)
{
- struct r600_query_hw *query = (struct r600_query_hw *)rquery;
struct r600_query_buffer *prev = query->buffer.previous;
- if (!r600_query_needs_begin(rquery->type)) {
- assert(0);
- return false;
- }
-
/* Discard the old query buffers. */
while (prev) {
struct r600_query_buffer *qbuf = prev;
FREE(qbuf);
}
- /* Obtain a new buffer if the current one can't be mapped without a stall. */
- if (r600_rings_is_buffer_referenced(rctx, query->buffer.buf->cs_buf, RADEON_USAGE_READWRITE) ||
- !rctx->ws->buffer_wait(query->buffer.buf->buf, 0, RADEON_USAGE_READWRITE)) {
- pipe_resource_reference((struct pipe_resource**)&query->buffer.buf, NULL);
- query->buffer.buf = r600_new_query_buffer(rctx, rquery->type);
+ if (query->flags & R600_QUERY_HW_FLAG_PREDICATE) {
+ /* Obtain a new buffer if the current one can't be mapped without a stall. */
+ if (r600_rings_is_buffer_referenced(rctx, query->buffer.buf->buf, RADEON_USAGE_READWRITE) ||
+ !rctx->ws->buffer_wait(query->buffer.buf->buf, 0, RADEON_USAGE_READWRITE)) {
+ pipe_resource_reference((struct pipe_resource**)&query->buffer.buf, NULL);
+ query->buffer.buf = r600_new_query_buffer(rctx, query);
+ } else {
+ query->ops->prepare_buffer(rctx, query, query->buffer.buf);
+ }
}
query->buffer.results_end = 0;
query->buffer.previous = NULL;
+}
+
+boolean r600_query_hw_begin(struct r600_common_context *rctx,
+ struct r600_query *rquery)
+{
+ struct r600_query_hw *query = (struct r600_query_hw *)rquery;
+
+ if (query->flags & R600_QUERY_HW_FLAG_NO_START) {
+ assert(0);
+ return false;
+ }
+
+ r600_query_hw_reset_buffers(rctx, query);
- r600_emit_query_begin(rctx, query);
+ r600_query_hw_emit_start(rctx, query);
- if (r600_is_timer_query(rquery->type))
+ if (query->flags & R600_QUERY_HW_FLAG_TIMER)
LIST_ADDTAIL(&query->list, &rctx->active_timer_queries);
else
LIST_ADDTAIL(&query->list, &rctx->active_nontimer_queries);
rquery->ops->end(rctx, rquery);
}
-static void r600_query_hw_end(struct r600_common_context *rctx,
+void r600_query_hw_end(struct r600_common_context *rctx,
struct r600_query *rquery)
{
struct r600_query_hw *query = (struct r600_query_hw *)rquery;
- r600_emit_query_end(rctx, query);
+ if (query->flags & R600_QUERY_HW_FLAG_NO_START)
+ r600_query_hw_reset_buffers(rctx, query);
+
+ r600_query_hw_emit_stop(rctx, query);
- if (r600_query_needs_begin(rquery->type))
+ if (!(query->flags & R600_QUERY_HW_FLAG_NO_START))
LIST_DELINIT(&query->list);
}
-static unsigned r600_query_read_result(char *map, unsigned start_index, unsigned end_index,
+static unsigned r600_query_read_result(void *map, unsigned start_index, unsigned end_index,
bool test_status_bit)
{
uint32_t *current_result = (uint32_t*)map;
return 0;
}
-static boolean r600_get_query_buffer_result(struct r600_common_context *ctx,
- struct r600_query_hw *query,
- struct r600_query_buffer *qbuf,
- boolean wait,
- union pipe_query_result *result)
+static void r600_query_hw_add_result(struct r600_common_context *ctx,
+ struct r600_query_hw *query,
+ void *buffer,
+ union pipe_query_result *result)
{
- unsigned results_base = 0;
- char *map;
-
- map = r600_buffer_map_sync_with_rings(ctx, qbuf->buf,
- PIPE_TRANSFER_READ |
- (wait ? 0 : PIPE_TRANSFER_DONTBLOCK));
- if (!map)
- return FALSE;
-
- /* count all results across all data blocks */
switch (query->b.type) {
- case PIPE_QUERY_OCCLUSION_COUNTER:
- while (results_base != qbuf->results_end) {
+ case PIPE_QUERY_OCCLUSION_COUNTER: {
+ unsigned results_base = 0;
+ while (results_base != query->result_size) {
result->u64 +=
- r600_query_read_result(map + results_base, 0, 2, true);
+ r600_query_read_result(buffer + results_base, 0, 2, true);
results_base += 16;
}
break;
- case PIPE_QUERY_OCCLUSION_PREDICATE:
- while (results_base != qbuf->results_end) {
+ }
+ case PIPE_QUERY_OCCLUSION_PREDICATE: {
+ unsigned results_base = 0;
+ while (results_base != query->result_size) {
result->b = result->b ||
- r600_query_read_result(map + results_base, 0, 2, true) != 0;
+ r600_query_read_result(buffer + results_base, 0, 2, true) != 0;
results_base += 16;
}
break;
+ }
case PIPE_QUERY_TIME_ELAPSED:
- while (results_base != qbuf->results_end) {
- result->u64 +=
- r600_query_read_result(map + results_base, 0, 2, false);
- results_base += query->result_size;
- }
+ result->u64 += r600_query_read_result(buffer, 0, 2, false);
break;
case PIPE_QUERY_TIMESTAMP:
{
- uint32_t *current_result = (uint32_t*)map;
+ uint32_t *current_result = (uint32_t*)buffer;
result->u64 = (uint64_t)current_result[0] |
(uint64_t)current_result[1] << 32;
break;
* u64 PrimitiveStorageNeeded;
* }
* We only need NumPrimitivesWritten here. */
- while (results_base != qbuf->results_end) {
- result->u64 +=
- r600_query_read_result(map + results_base, 2, 6, true);
- results_base += query->result_size;
- }
+ result->u64 += r600_query_read_result(buffer, 2, 6, true);
break;
case PIPE_QUERY_PRIMITIVES_GENERATED:
/* Here we read PrimitiveStorageNeeded. */
- while (results_base != qbuf->results_end) {
- result->u64 +=
- r600_query_read_result(map + results_base, 0, 4, true);
- results_base += query->result_size;
- }
+ result->u64 += r600_query_read_result(buffer, 0, 4, true);
break;
case PIPE_QUERY_SO_STATISTICS:
- while (results_base != qbuf->results_end) {
- result->so_statistics.num_primitives_written +=
- r600_query_read_result(map + results_base, 2, 6, true);
- result->so_statistics.primitives_storage_needed +=
- r600_query_read_result(map + results_base, 0, 4, true);
- results_base += query->result_size;
- }
+ result->so_statistics.num_primitives_written +=
+ r600_query_read_result(buffer, 2, 6, true);
+ result->so_statistics.primitives_storage_needed +=
+ r600_query_read_result(buffer, 0, 4, true);
break;
case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
- while (results_base != qbuf->results_end) {
- result->b = result->b ||
- r600_query_read_result(map + results_base, 2, 6, true) !=
- r600_query_read_result(map + results_base, 0, 4, true);
- results_base += query->result_size;
- }
+ result->b = result->b ||
+ r600_query_read_result(buffer, 2, 6, true) !=
+ r600_query_read_result(buffer, 0, 4, true);
break;
case PIPE_QUERY_PIPELINE_STATISTICS:
if (ctx->chip_class >= EVERGREEN) {
- while (results_base != qbuf->results_end) {
- result->pipeline_statistics.ps_invocations +=
- r600_query_read_result(map + results_base, 0, 22, false);
- result->pipeline_statistics.c_primitives +=
- r600_query_read_result(map + results_base, 2, 24, false);
- result->pipeline_statistics.c_invocations +=
- r600_query_read_result(map + results_base, 4, 26, false);
- result->pipeline_statistics.vs_invocations +=
- r600_query_read_result(map + results_base, 6, 28, false);
- result->pipeline_statistics.gs_invocations +=
- r600_query_read_result(map + results_base, 8, 30, false);
- result->pipeline_statistics.gs_primitives +=
- r600_query_read_result(map + results_base, 10, 32, false);
- result->pipeline_statistics.ia_primitives +=
- r600_query_read_result(map + results_base, 12, 34, false);
- result->pipeline_statistics.ia_vertices +=
- r600_query_read_result(map + results_base, 14, 36, false);
- result->pipeline_statistics.hs_invocations +=
- r600_query_read_result(map + results_base, 16, 38, false);
- result->pipeline_statistics.ds_invocations +=
- r600_query_read_result(map + results_base, 18, 40, false);
- result->pipeline_statistics.cs_invocations +=
- r600_query_read_result(map + results_base, 20, 42, false);
- results_base += query->result_size;
- }
+ result->pipeline_statistics.ps_invocations +=
+ r600_query_read_result(buffer, 0, 22, false);
+ result->pipeline_statistics.c_primitives +=
+ r600_query_read_result(buffer, 2, 24, false);
+ result->pipeline_statistics.c_invocations +=
+ r600_query_read_result(buffer, 4, 26, false);
+ result->pipeline_statistics.vs_invocations +=
+ r600_query_read_result(buffer, 6, 28, false);
+ result->pipeline_statistics.gs_invocations +=
+ r600_query_read_result(buffer, 8, 30, false);
+ result->pipeline_statistics.gs_primitives +=
+ r600_query_read_result(buffer, 10, 32, false);
+ result->pipeline_statistics.ia_primitives +=
+ r600_query_read_result(buffer, 12, 34, false);
+ result->pipeline_statistics.ia_vertices +=
+ r600_query_read_result(buffer, 14, 36, false);
+ result->pipeline_statistics.hs_invocations +=
+ r600_query_read_result(buffer, 16, 38, false);
+ result->pipeline_statistics.ds_invocations +=
+ r600_query_read_result(buffer, 18, 40, false);
+ result->pipeline_statistics.cs_invocations +=
+ r600_query_read_result(buffer, 20, 42, false);
} else {
- while (results_base != qbuf->results_end) {
- result->pipeline_statistics.ps_invocations +=
- r600_query_read_result(map + results_base, 0, 16, false);
- result->pipeline_statistics.c_primitives +=
- r600_query_read_result(map + results_base, 2, 18, false);
- result->pipeline_statistics.c_invocations +=
- r600_query_read_result(map + results_base, 4, 20, false);
- result->pipeline_statistics.vs_invocations +=
- r600_query_read_result(map + results_base, 6, 22, false);
- result->pipeline_statistics.gs_invocations +=
- r600_query_read_result(map + results_base, 8, 24, false);
- result->pipeline_statistics.gs_primitives +=
- r600_query_read_result(map + results_base, 10, 26, false);
- result->pipeline_statistics.ia_primitives +=
- r600_query_read_result(map + results_base, 12, 28, false);
- result->pipeline_statistics.ia_vertices +=
- r600_query_read_result(map + results_base, 14, 30, false);
- results_base += query->result_size;
- }
+ result->pipeline_statistics.ps_invocations +=
+ r600_query_read_result(buffer, 0, 16, false);
+ result->pipeline_statistics.c_primitives +=
+ r600_query_read_result(buffer, 2, 18, false);
+ result->pipeline_statistics.c_invocations +=
+ r600_query_read_result(buffer, 4, 20, false);
+ result->pipeline_statistics.vs_invocations +=
+ r600_query_read_result(buffer, 6, 22, false);
+ result->pipeline_statistics.gs_invocations +=
+ r600_query_read_result(buffer, 8, 24, false);
+ result->pipeline_statistics.gs_primitives +=
+ r600_query_read_result(buffer, 10, 26, false);
+ result->pipeline_statistics.ia_primitives +=
+ r600_query_read_result(buffer, 12, 28, false);
+ result->pipeline_statistics.ia_vertices +=
+ r600_query_read_result(buffer, 14, 30, false);
}
#if 0 /* for testing */
printf("Pipeline stats: IA verts=%llu, IA prims=%llu, VS=%llu, HS=%llu, "
default:
assert(0);
}
-
- return TRUE;
}
static boolean r600_get_query_result(struct pipe_context *ctx,
return rquery->ops->get_result(rctx, rquery, wait, result);
}
-static boolean r600_query_hw_get_result(struct r600_common_context *rctx,
- struct r600_query *rquery,
- boolean wait, union pipe_query_result *result)
+static void r600_query_hw_clear_result(struct r600_query_hw *query,
+ union pipe_query_result *result)
+{
+ util_query_clear_result(result, query->b.type);
+}
+
+boolean r600_query_hw_get_result(struct r600_common_context *rctx,
+ struct r600_query *rquery,
+ boolean wait, union pipe_query_result *result)
{
struct r600_query_hw *query = (struct r600_query_hw *)rquery;
struct r600_query_buffer *qbuf;
- util_query_clear_result(result, rquery->type);
+ query->ops->clear_result(query, result);
for (qbuf = &query->buffer; qbuf; qbuf = qbuf->previous) {
- if (!r600_get_query_buffer_result(rctx, query, qbuf, wait, result)) {
+ unsigned results_base = 0;
+ void *map;
+
+ map = r600_buffer_map_sync_with_rings(rctx, qbuf->buf,
+ PIPE_TRANSFER_READ |
+ (wait ? 0 : PIPE_TRANSFER_DONTBLOCK));
+ if (!map)
return FALSE;
+
+ while (results_base != qbuf->results_end) {
+ query->ops->add_result(rctx, query, map + results_base,
+ result);
+ results_base += query->result_size;
}
}
/* Convert the time to expected units. */
if (rquery->type == PIPE_QUERY_TIME_ELAPSED ||
rquery->type == PIPE_QUERY_TIMESTAMP) {
- result->u64 = (1000000 * result->u64) / rctx->screen->info.r600_clock_crystal_freq;
+ result->u64 = (1000000 * result->u64) / rctx->screen->info.clock_crystal_freq;
}
return TRUE;
}
struct r600_query_hw *query;
LIST_FOR_EACH_ENTRY(query, query_list, list) {
- r600_emit_query_end(ctx, query);
+ r600_query_hw_emit_stop(ctx, query);
}
assert(*num_cs_dw_queries_suspend == 0);
}
LIST_FOR_EACH_ENTRY(query, query_list, list) {
/* begin + end */
- num_dw += query->num_cs_dw * 2;
+ num_dw += query->num_cs_dw_begin + query->num_cs_dw_end;
/* Workaround for the fact that
* num_cs_dw_nontimer_queries_suspend is incremented for every
* resumed query, which raises the bar in need_cs_space for
* queries about to be resumed.
*/
- num_dw += query->num_cs_dw;
+ num_dw += query->num_cs_dw_end;
}
/* primitives generated query */
num_dw += ctx->streamout.enable_atom.num_dw;
ctx->need_gfx_cs_space(&ctx->b, num_cs_dw, TRUE);
LIST_FOR_EACH_ENTRY(query, query_list, list) {
- r600_emit_query_begin(ctx, query);
+ r600_query_hw_emit_start(ctx, query);
}
}
struct radeon_winsys_cs *cs = ctx->gfx.cs;
struct r600_resource *buffer;
uint32_t *results;
- unsigned num_backends = ctx->screen->info.r600_num_backends;
+ unsigned num_backends = ctx->screen->info.num_render_backends;
unsigned i, mask = 0;
/* if backend_map query is supported by the kernel */
- if (ctx->screen->info.r600_backend_map_valid) {
- unsigned num_tile_pipes = ctx->screen->info.r600_num_tile_pipes;
- unsigned backend_map = ctx->screen->info.r600_backend_map;
+ if (ctx->screen->info.r600_gb_backend_map_valid) {
+ unsigned num_tile_pipes = ctx->screen->info.num_tile_pipes;
+ unsigned backend_map = ctx->screen->info.r600_gb_backend_map;
unsigned item_width, item_mask;
if (ctx->chip_class >= EVERGREEN) {
return;
}
-#define X(name_, query_type_, type_, result_type_) \
+#define XFULL(name_, query_type_, type_, result_type_, group_id_) \
{ \
.name = name_, \
.query_type = R600_QUERY_##query_type_, \
.type = PIPE_DRIVER_QUERY_TYPE_##type_, \
.result_type = PIPE_DRIVER_QUERY_RESULT_TYPE_##result_type_, \
- .group_id = ~(unsigned)0 \
+ .group_id = group_id_ \
}
+#define X(name_, query_type_, type_, result_type_) \
+ XFULL(name_, query_type_, type_, result_type_, ~(unsigned)0)
+
+#define XG(group_, name_, query_type_, type_, result_type_) \
+ XFULL(name_, query_type_, type_, result_type_, R600_QUERY_GROUP_##group_)
+
static struct pipe_driver_query_info r600_driver_query_list[] = {
X("num-compilations", NUM_COMPILATIONS, UINT64, CUMULATIVE),
X("num-shaders-created", NUM_SHADERS_CREATED, UINT64, CUMULATIVE),
X("num-bytes-moved", NUM_BYTES_MOVED, BYTES, CUMULATIVE),
X("VRAM-usage", VRAM_USAGE, BYTES, AVERAGE),
X("GTT-usage", GTT_USAGE, BYTES, AVERAGE),
+
+ /* GPIN queries are for the benefit of old versions of GPUPerfStudio,
+ * which use it as a fallback path to detect the GPU type.
+ *
+ * Note: The names of these queries are significant for GPUPerfStudio
+ * (and possibly their order as well). */
+ XG(GPIN, "GPIN_000", GPIN_ASIC_ID, UINT, AVERAGE),
+ XG(GPIN, "GPIN_001", GPIN_NUM_SIMD, UINT, AVERAGE),
+ XG(GPIN, "GPIN_002", GPIN_NUM_RB, UINT, AVERAGE),
+ XG(GPIN, "GPIN_003", GPIN_NUM_SPI, UINT, AVERAGE),
+ XG(GPIN, "GPIN_004", GPIN_NUM_SE, UINT, AVERAGE),
+
+ /* The following queries must be at the end of the list because their
+ * availability is adjusted dynamically based on the DRM version. */
X("GPU-load", GPU_LOAD, UINT64, AVERAGE),
X("temperature", GPU_TEMPERATURE, UINT64, AVERAGE),
X("shader-clock", CURRENT_GPU_SCLK, HZ, AVERAGE),
};
#undef X
+#undef XG
+#undef XFULL
static unsigned r600_get_num_queries(struct r600_common_screen *rscreen)
{
struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
unsigned num_queries = r600_get_num_queries(rscreen);
- if (!info)
- return num_queries;
+ if (!info) {
+ unsigned num_perfcounters =
+ r600_get_perfcounter_info(rscreen, 0, NULL);
+
+ return num_queries + num_perfcounters;
+ }
if (index >= num_queries)
- return 0;
+ return r600_get_perfcounter_info(rscreen, index - num_queries, info);
*info = r600_driver_query_list[index];
break;
}
+ if (info->group_id != ~(unsigned)0 && rscreen->perfcounters)
+ info->group_id += rscreen->perfcounters->num_groups;
+
+ return 1;
+}
+
+/* Note: Unfortunately, GPUPerfStudio hardcodes the order of hardware
+ * performance counter groups, so be careful when changing this and related
+ * functions.
+ */
+static int r600_get_driver_query_group_info(struct pipe_screen *screen,
+ unsigned index,
+ struct pipe_driver_query_group_info *info)
+{
+ struct r600_common_screen *rscreen = (struct r600_common_screen *)screen;
+ unsigned num_pc_groups = 0;
+
+ if (rscreen->perfcounters)
+ num_pc_groups = rscreen->perfcounters->num_groups;
+
+ if (!info)
+ return num_pc_groups + R600_NUM_SW_QUERY_GROUPS;
+
+ if (index < num_pc_groups)
+ return r600_get_perfcounter_group_info(rscreen, index, info);
+
+ index -= num_pc_groups;
+ if (index >= R600_NUM_SW_QUERY_GROUPS)
+ return 0;
+
+ info->name = "GPIN";
+ info->max_active_queries = 5;
+ info->num_queries = 5;
return 1;
}
void r600_query_init(struct r600_common_context *rctx)
{
rctx->b.create_query = r600_create_query;
+ rctx->b.create_batch_query = r600_create_batch_query;
rctx->b.destroy_query = r600_destroy_query;
rctx->b.begin_query = r600_begin_query;
rctx->b.end_query = r600_end_query;
rctx->b.get_query_result = r600_get_query_result;
rctx->render_cond_atom.emit = r600_emit_query_predication;
- if (((struct r600_common_screen*)rctx->b.screen)->info.r600_num_backends > 0)
+ if (((struct r600_common_screen*)rctx->b.screen)->info.num_render_backends > 0)
rctx->b.render_condition = r600_render_condition;
LIST_INITHEAD(&rctx->active_nontimer_queries);
void r600_init_screen_query_functions(struct r600_common_screen *rscreen)
{
rscreen->b.get_driver_query_info = r600_get_driver_query_info;
+ rscreen->b.get_driver_query_group_info = r600_get_driver_query_group_info;
}