#include "brw_defines.h"
#include "brw_state.h"
#include "intel_batchbuffer.h"
+#include "intel_buffer_objects.h"
#include "intel_reg.h"
-/**
- * Emit PIPE_CONTROLs to write the current GPU timestamp into a buffer.
- */
+static inline void
+set_query_availability(struct brw_context *brw, struct brw_query_object *query,
+ bool available)
+{
+ /* For platforms that support ARB_query_buffer_object, we write the
+ * query availability for "pipelined" queries.
+ *
+ * Most counter snapshots are written by the command streamer, by
+ * doing a CS stall and then MI_STORE_REGISTER_MEM. For these
+ * counters, the CS stall guarantees that the results will be
+ * available when subsequent CS commands run. So we don't need to
+ * do any additional tracking.
+ *
+ * Other counters (occlusion queries and timestamp) are written by
+ * PIPE_CONTROL, without a CS stall. This means that we can't be
+ * sure whether the writes have landed yet or not. Performing a
+ * PIPE_CONTROL with an immediate write will synchronize with
+ * those earlier writes, so we write 1 when the value has landed.
+ */
+ if (brw->ctx.Extensions.ARB_query_buffer_object &&
+ brw_is_query_pipelined(query)) {
+ brw_emit_pipe_control_write(brw,
+ PIPE_CONTROL_WRITE_IMMEDIATE,
+ query->bo, 2 * sizeof(uint64_t),
+ available, 0);
+ }
+}
+
static void
-write_timestamp(struct intel_context *intel, drm_intel_bo *query_bo, int idx)
+write_primitives_generated(struct brw_context *brw,
+ drm_intel_bo *query_bo, int stream, int idx)
{
- /* Emit workaround flushes: */
- if (intel->gen == 6) {
- /* The timestamp write below is a non-zero post-sync op, which on
- * Gen6 necessitates a CS stall. CS stalls need stall at scoreboard
- * set. See the comments for intel_emit_post_sync_nonzero_flush().
- */
- BEGIN_BATCH(4);
- OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2));
- OUT_BATCH(PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD);
- OUT_BATCH(0);
- OUT_BATCH(0);
- ADVANCE_BATCH();
+ brw_emit_mi_flush(brw);
+
+ if (brw->gen >= 7 && stream > 0) {
+ brw_store_register_mem64(brw, query_bo,
+ GEN7_SO_PRIM_STORAGE_NEEDED(stream),
+ idx * sizeof(uint64_t));
+ } else {
+ brw_store_register_mem64(brw, query_bo, CL_INVOCATION_COUNT,
+ idx * sizeof(uint64_t));
}
+}
- BEGIN_BATCH(5);
- OUT_BATCH(_3DSTATE_PIPE_CONTROL | (5 - 2));
- OUT_BATCH(PIPE_CONTROL_WRITE_TIMESTAMP);
- OUT_RELOC(query_bo,
- I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
- PIPE_CONTROL_GLOBAL_GTT_WRITE |
- idx * sizeof(uint64_t));
- OUT_BATCH(0);
- OUT_BATCH(0);
- ADVANCE_BATCH();
+static void
+write_xfb_primitives_written(struct brw_context *brw,
+ drm_intel_bo *bo, int stream, int idx)
+{
+ brw_emit_mi_flush(brw);
+
+ if (brw->gen >= 7) {
+ brw_store_register_mem64(brw, bo, GEN7_SO_NUM_PRIMS_WRITTEN(stream),
+ idx * sizeof(uint64_t));
+ } else {
+ brw_store_register_mem64(brw, bo, GEN6_SO_NUM_PRIMS_WRITTEN,
+ idx * sizeof(uint64_t));
+ }
+}
+
+static inline const int
+pipeline_target_to_index(int target)
+{
+ if (target == GL_GEOMETRY_SHADER_INVOCATIONS)
+ return MAX_PIPELINE_STATISTICS - 1;
+ else
+ return target - GL_VERTICES_SUBMITTED_ARB;
}
-/**
- * Emit PIPE_CONTROLs to write the PS_DEPTH_COUNT register into a buffer.
- */
static void
-write_depth_count(struct intel_context *intel, drm_intel_bo *query_bo, int idx)
+emit_pipeline_stat(struct brw_context *brw, drm_intel_bo *bo,
+ int stream, int target, int idx)
{
- /* Emit Sandybridge workaround flush: */
- if (intel->gen == 6)
- intel_emit_post_sync_nonzero_flush(intel);
-
- BEGIN_BATCH(5);
- OUT_BATCH(_3DSTATE_PIPE_CONTROL | (5 - 2));
- OUT_BATCH(PIPE_CONTROL_DEPTH_STALL |
- PIPE_CONTROL_WRITE_DEPTH_COUNT);
- OUT_RELOC(query_bo,
- I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
- PIPE_CONTROL_GLOBAL_GTT_WRITE |
- (idx * sizeof(uint64_t)));
- OUT_BATCH(0);
- OUT_BATCH(0);
- ADVANCE_BATCH();
+ /* One source of confusion is the tessellation shader statistics. The
+ * hardware has no statistics specific to the TE unit. Ideally we could have
+ * the HS primitives for TESS_CONTROL_SHADER_PATCHES_ARB, and the DS
+ * invocations as the register for TESS_CONTROL_SHADER_PATCHES_ARB.
+ * Unfortunately we don't have HS primitives, we only have HS invocations.
+ */
+
+ /* Everything except GEOMETRY_SHADER_INVOCATIONS can be kept in a simple
+ * lookup table
+ */
+ static const uint32_t target_to_register[] = {
+ IA_VERTICES_COUNT, /* VERTICES_SUBMITTED */
+ IA_PRIMITIVES_COUNT, /* PRIMITIVES_SUBMITTED */
+ VS_INVOCATION_COUNT, /* VERTEX_SHADER_INVOCATIONS */
+ HS_INVOCATION_COUNT, /* TESS_CONTROL_SHADER_PATCHES */
+ DS_INVOCATION_COUNT, /* TESS_EVALUATION_SHADER_INVOCATIONS */
+ GS_PRIMITIVES_COUNT, /* GEOMETRY_SHADER_PRIMITIVES_EMITTED */
+ PS_INVOCATION_COUNT, /* FRAGMENT_SHADER_INVOCATIONS */
+ CS_INVOCATION_COUNT, /* COMPUTE_SHADER_INVOCATIONS */
+ CL_INVOCATION_COUNT, /* CLIPPING_INPUT_PRIMITIVES */
+ CL_PRIMITIVES_COUNT, /* CLIPPING_OUTPUT_PRIMITIVES */
+ GS_INVOCATION_COUNT /* This one is special... */
+ };
+ STATIC_ASSERT(ARRAY_SIZE(target_to_register) == MAX_PIPELINE_STATISTICS);
+ uint32_t reg = target_to_register[pipeline_target_to_index(target)];
+ /* Gen6 GS code counts full primitives, that is, it won't count individual
+ * triangles in a triangle strip. Use CL_INVOCATION_COUNT for that.
+ */
+ if (brw->gen == 6 && target == GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB)
+ reg = CL_INVOCATION_COUNT;
+ assert(reg != 0);
+
+ /* Emit a flush to make sure various parts of the pipeline are complete and
+ * we get an accurate value
+ */
+ brw_emit_mi_flush(brw);
+
+ brw_store_register_mem64(brw, bo, reg, idx * sizeof(uint64_t));
}
+
/**
* Wait on the query object's BO and calculate the final result.
*/
gen6_queryobj_get_results(struct gl_context *ctx,
struct brw_query_object *query)
{
- struct intel_context *intel = intel_context(ctx);
+ struct brw_context *brw = brw_context(ctx);
if (query->bo == NULL)
return;
- /* If the application has requested the query result, but this batch is
- * still contributing to it, flush it now so the results will be present
- * when mapped.
- */
- if (drm_intel_bo_references(intel->batch.bo, query->bo))
- intel_batchbuffer_flush(intel);
-
- if (unlikely(intel->perf_debug)) {
- if (drm_intel_bo_busy(query->bo)) {
- perf_debug("Stalling on the GPU waiting for a query object.\n");
- }
- }
-
- drm_intel_bo_map(query->bo, false);
+ brw_bo_map(brw, query->bo, false, "query object");
uint64_t *results = query->bo->virtual;
switch (query->Base.Target) {
case GL_TIME_ELAPSED:
case GL_PRIMITIVES_GENERATED:
case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
- /* We don't actually query the hardware for this value, so query->bo
- * should always be NULL and execution should never reach here.
+ case GL_VERTICES_SUBMITTED_ARB:
+ case GL_PRIMITIVES_SUBMITTED_ARB:
+ case GL_VERTEX_SHADER_INVOCATIONS_ARB:
+ case GL_GEOMETRY_SHADER_INVOCATIONS:
+ case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB:
+ case GL_CLIPPING_INPUT_PRIMITIVES_ARB:
+ case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB:
+ case GL_COMPUTE_SHADER_INVOCATIONS_ARB:
+ case GL_TESS_CONTROL_SHADER_PATCHES_ARB:
+ case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB:
+ query->Base.Result = results[1] - results[0];
+ break;
+
+ case GL_FRAGMENT_SHADER_INVOCATIONS_ARB:
+ query->Base.Result = (results[1] - results[0]);
+ /* Implement the "WaDividePSInvocationCountBy4:HSW,BDW" workaround:
+ * "Invocation counter is 4 times actual. WA: SW to divide HW reported
+ * PS Invocations value by 4."
+ *
+ * Prior to Haswell, invocation count was counted by the WM, and it
+ * buggily counted invocations in units of subspans (2x2 unit). To get the
+ * correct value, the CS multiplied this by 4. With HSW the logic moved,
+ * and correctly emitted the number of pixel shader invocations, but,
+ * whomever forgot to undo the multiply by 4.
*/
- assert(!"Unreachable");
+ if (brw->gen == 8 || brw->is_haswell)
+ query->Base.Result /= 4;
break;
default:
- assert(!"Unrecognized query target in brw_queryobj_get_results()");
- break;
+ unreachable("Unrecognized query target in brw_queryobj_get_results()");
}
drm_intel_bo_unmap(query->bo);
*/
drm_intel_bo_unreference(query->bo);
query->bo = NULL;
+
+ query->Base.Ready = true;
}
/**
gen6_begin_query(struct gl_context *ctx, struct gl_query_object *q)
{
struct brw_context *brw = brw_context(ctx);
- struct intel_context *intel = intel_context(ctx);
struct brw_query_object *query = (struct brw_query_object *)q;
+ /* Since we're starting a new query, we need to throw away old results. */
+ drm_intel_bo_unreference(query->bo);
+ query->bo = drm_intel_bo_alloc(brw->bufmgr, "query results", 4096, 4096);
+
+ /* For ARB_query_buffer_object: The result is not available */
+ set_query_availability(brw, query, false);
+
switch (query->Base.Target) {
case GL_TIME_ELAPSED:
/* For timestamp queries, we record the starting time right away so that
* obtain the time elapsed. Notably, this includes time elapsed while
* the system was doing other work, such as running other applications.
*/
- drm_intel_bo_unreference(query->bo);
- query->bo = drm_intel_bo_alloc(intel->bufmgr, "timer query", 4096, 4096);
- write_timestamp(intel, query->bo, 0);
+ brw_write_timestamp(brw, query->bo, 0);
break;
case GL_ANY_SAMPLES_PASSED:
case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
case GL_SAMPLES_PASSED_ARB:
- /* Since we're starting a new query, we need to be sure to throw away
- * any previous occlusion query results.
- */
- drm_intel_bo_unreference(query->bo);
- query->bo = drm_intel_bo_alloc(intel->bufmgr, "occl. query", 4096, 4096);
- write_depth_count(intel, query->bo, 0);
+ brw_write_depth_count(brw, query->bo, 0);
break;
case GL_PRIMITIVES_GENERATED:
- /* We don't actually query the hardware for this value; we keep track of
- * it a software counter. So just reset the counter.
- */
- brw->sol.primitives_generated = 0;
- brw->sol.counting_primitives_generated = true;
+ write_primitives_generated(brw, query->bo, query->Base.Stream, 0);
break;
case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
- /* We don't actually query the hardware for this value; we keep track of
- * it a software counter. So just reset the counter.
- */
- brw->sol.primitives_written = 0;
- brw->sol.counting_primitives_written = true;
+ write_xfb_primitives_written(brw, query->bo, query->Base.Stream, 0);
break;
- default:
- assert(!"Unrecognized query target in brw_begin_query()");
+ case GL_VERTICES_SUBMITTED_ARB:
+ case GL_PRIMITIVES_SUBMITTED_ARB:
+ case GL_VERTEX_SHADER_INVOCATIONS_ARB:
+ case GL_GEOMETRY_SHADER_INVOCATIONS:
+ case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB:
+ case GL_FRAGMENT_SHADER_INVOCATIONS_ARB:
+ case GL_CLIPPING_INPUT_PRIMITIVES_ARB:
+ case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB:
+ case GL_COMPUTE_SHADER_INVOCATIONS_ARB:
+ case GL_TESS_CONTROL_SHADER_PATCHES_ARB:
+ case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB:
+ emit_pipeline_stat(brw, query->bo, query->Base.Stream, query->Base.Target, 0);
break;
+
+ default:
+ unreachable("Unrecognized query target in brw_begin_query()");
}
}
gen6_end_query(struct gl_context *ctx, struct gl_query_object *q)
{
struct brw_context *brw = brw_context(ctx);
- struct intel_context *intel = intel_context(ctx);
struct brw_query_object *query = (struct brw_query_object *)q;
switch (query->Base.Target) {
case GL_TIME_ELAPSED:
- write_timestamp(intel, query->bo, 1);
+ brw_write_timestamp(brw, query->bo, 1);
break;
case GL_ANY_SAMPLES_PASSED:
case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
case GL_SAMPLES_PASSED_ARB:
- write_depth_count(intel, query->bo, 1);
+ brw_write_depth_count(brw, query->bo, 1);
break;
case GL_PRIMITIVES_GENERATED:
- /* We don't actually query the hardware for this value; we keep track of
- * it in a software counter. So just read the counter and store it in
- * the query object.
- */
- query->Base.Result = brw->sol.primitives_generated;
- brw->sol.counting_primitives_generated = false;
+ write_primitives_generated(brw, query->bo, query->Base.Stream, 1);
break;
case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
- /* We don't actually query the hardware for this value; we keep track of
- * it in a software counter. So just read the counter and store it in
- * the query object.
- */
- query->Base.Result = brw->sol.primitives_written;
- brw->sol.counting_primitives_written = false;
+ write_xfb_primitives_written(brw, query->bo, query->Base.Stream, 1);
break;
- default:
- assert(!"Unrecognized query target in brw_end_query()");
+ case GL_VERTICES_SUBMITTED_ARB:
+ case GL_PRIMITIVES_SUBMITTED_ARB:
+ case GL_VERTEX_SHADER_INVOCATIONS_ARB:
+ case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB:
+ case GL_FRAGMENT_SHADER_INVOCATIONS_ARB:
+ case GL_COMPUTE_SHADER_INVOCATIONS_ARB:
+ case GL_CLIPPING_INPUT_PRIMITIVES_ARB:
+ case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB:
+ case GL_GEOMETRY_SHADER_INVOCATIONS:
+ case GL_TESS_CONTROL_SHADER_PATCHES_ARB:
+ case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB:
+ emit_pipeline_stat(brw, query->bo,
+ query->Base.Stream, query->Base.Target, 1);
break;
+
+ default:
+ unreachable("Unrecognized query target in brw_end_query()");
}
+
+ /* The current batch contains the commands to handle EndQuery(),
+ * but they won't actually execute until it is flushed.
+ */
+ query->flushed = false;
+
+ /* For ARB_query_buffer_object: The result is now available */
+ set_query_availability(brw, query, true);
+}
+
+/**
+ * Flush the batch if it still references the query object BO.
+ */
+static void
+flush_batch_if_needed(struct brw_context *brw, struct brw_query_object *query)
+{
+ /* If the batch doesn't reference the BO, it must have been flushed
+ * (for example, due to being full). Record that it's been flushed.
+ */
+ query->flushed = query->flushed ||
+ !drm_intel_bo_references(brw->batch.bo, query->bo);
+
+ if (!query->flushed)
+ intel_batchbuffer_flush(brw);
}
/**
*/
static void gen6_wait_query(struct gl_context *ctx, struct gl_query_object *q)
{
+ struct brw_context *brw = brw_context(ctx);
struct brw_query_object *query = (struct brw_query_object *)q;
+ /* If the application has requested the query result, but this batch is
+ * still contributing to it, flush it now to finish that work so the
+ * result will become available (eventually).
+ */
+ flush_batch_if_needed(brw, query);
+
gen6_queryobj_get_results(ctx, query);
- query->Base.Ready = true;
}
/**
*/
static void gen6_check_query(struct gl_context *ctx, struct gl_query_object *q)
{
- struct intel_context *intel = intel_context(ctx);
+ struct brw_context *brw = brw_context(ctx);
struct brw_query_object *query = (struct brw_query_object *)q;
+ /* If query->bo is NULL, we've already gathered the results - this is a
+ * redundant CheckQuery call. Ignore it.
+ */
+ if (query->bo == NULL)
+ return;
+
/* From the GL_ARB_occlusion_query spec:
*
* "Instead of allowing for an infinite loop, performing a
* not ready yet on the first time it is queried. This ensures that
* the async query will return true in finite time.
*/
- if (query->bo && drm_intel_bo_references(intel->batch.bo, query->bo))
- intel_batchbuffer_flush(intel);
+ flush_batch_if_needed(brw, query);
- if (query->bo == NULL || !drm_intel_bo_busy(query->bo)) {
+ if (!drm_intel_bo_busy(query->bo)) {
gen6_queryobj_get_results(ctx, query);
- query->Base.Ready = true;
}
}
+static void
+gen6_query_counter(struct gl_context *ctx, struct gl_query_object *q)
+{
+ struct brw_context *brw = brw_context(ctx);
+ struct brw_query_object *query = (struct brw_query_object *)q;
+ brw_query_counter(ctx, q);
+ set_query_availability(brw, query, true);
+}
+
/* Initialize Gen6+-specific query object functions. */
void gen6_init_queryobj_functions(struct dd_function_table *functions)
{
functions->EndQuery = gen6_end_query;
functions->CheckQuery = gen6_check_query;
functions->WaitQuery = gen6_wait_query;
+ functions->QueryCounter = gen6_query_counter;
}