#include "brw_defines.h"
#include "brw_state.h"
#include "intel_batchbuffer.h"
-#include "intel_reg.h"
-/**
- * Emit PIPE_CONTROLs to write the current GPU timestamp into a buffer.
+uint64_t
+brw_timebase_scale(struct brw_context *brw, uint64_t gpu_timestamp)
+{
+ const struct gen_device_info *devinfo = &brw->screen->devinfo;
+
+ return (1000000000ull * gpu_timestamp) / devinfo->timestamp_frequency;
+}
+
+/* As best we know currently, the Gen HW timestamps are 36bits across
+ * all platforms, which we need to account for when calculating a
+ * delta to measure elapsed time.
+ *
+ * The timestamps read via glGetTimestamp() / brw_get_timestamp() sometimes
+ * only have 32bits due to a kernel bug and so in that case we make sure to
+ * treat all raw timestamps as 32bits so they overflow consistently and remain
+ * comparable. (Note: the timestamps being passed here are not from the kernel
+ * so we don't need to be taking the upper 32bits in this buggy kernel case we
+ * are just clipping to 32bits here for consistency.)
*/
-static void
-write_timestamp(struct intel_context *intel, drm_intel_bo *query_bo, int idx)
+uint64_t
+brw_raw_timestamp_delta(struct brw_context *brw, uint64_t time0, uint64_t time1)
{
- if (intel->gen >= 6) {
- /* 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();
+ if (brw->screen->hw_has_timestamp == 2) {
+ /* Kernel clips timestamps to 32bits in this case, so we also clip
+ * PIPE_CONTROL timestamps for consistency.
+ */
+ return (uint32_t)time1 - (uint32_t)time0;
+ } else {
+ if (time0 > time1) {
+ return (1ULL << 36) + time1 - time0;
+ } else {
+ return time1 - time0;
}
+ }
+}
- 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();
- } else {
- BEGIN_BATCH(4);
- OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2) |
- 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();
+/**
+ * Emit PIPE_CONTROLs to write the current GPU timestamp into a buffer.
+ */
+void
+brw_write_timestamp(struct brw_context *brw, struct brw_bo *query_bo, int idx)
+{
+ if (brw->gen == 6) {
+ /* Emit Sandybridge workaround flush: */
+ brw_emit_pipe_control_flush(brw,
+ PIPE_CONTROL_CS_STALL |
+ PIPE_CONTROL_STALL_AT_SCOREBOARD);
}
+
+ uint32_t flags = PIPE_CONTROL_WRITE_TIMESTAMP;
+
+ if (brw->gen == 9 && brw->gt == 4)
+ flags |= PIPE_CONTROL_CS_STALL;
+
+ brw_emit_pipe_control_write(brw, flags,
+ query_bo, idx * sizeof(uint64_t), 0);
}
/**
* 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)
+void
+brw_write_depth_count(struct brw_context *brw, struct brw_bo *query_bo, int idx)
{
- if (intel->gen >= 6) {
- /* 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();
- } else {
- BEGIN_BATCH(4);
- OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2) |
- PIPE_CONTROL_DEPTH_STALL |
- PIPE_CONTROL_WRITE_DEPTH_COUNT);
- /* This object could be mapped cacheable, but we don't have an exposed
- * mechanism to support that. Since it's going uncached, tell GEM that
- * we're writing to it. The usual clflush should be all that's required
- * to pick up the results.
+ uint32_t flags = PIPE_CONTROL_WRITE_DEPTH_COUNT | PIPE_CONTROL_DEPTH_STALL;
+
+ if (brw->gen == 9 && brw->gt == 4)
+ flags |= PIPE_CONTROL_CS_STALL;
+
+ if (brw->gen >= 10) {
+ /* "Driver must program PIPE_CONTROL with only Depth Stall Enable bit set
+ * prior to programming a PIPE_CONTROL with Write PS Depth Count Post sync
+ * operation."
*/
- 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();
+ brw_emit_pipe_control_flush(brw, PIPE_CONTROL_DEPTH_STALL);
}
+
+ brw_emit_pipe_control_write(brw, flags,
+ query_bo, idx * sizeof(uint64_t), 0);
}
/**
brw_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);
int i;
uint64_t *results;
+ assert(brw->gen < 6);
+
if (query->bo == NULL)
return;
* 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 (brw_batch_references(&brw->batch, query->bo))
+ intel_batchbuffer_flush(brw);
- if (unlikely(intel->perf_debug)) {
- if (drm_intel_bo_busy(query->bo)) {
+ if (unlikely(brw->perf_debug)) {
+ if (brw_bo_busy(query->bo)) {
perf_debug("Stalling on the GPU waiting for a query object.\n");
}
}
- drm_intel_bo_map(query->bo, false);
- results = query->bo->virtual;
+ results = brw_bo_map(brw, query->bo, MAP_READ);
switch (query->Base.Target) {
case GL_TIME_ELAPSED_EXT:
/* The query BO contains the starting and ending timestamps.
* Subtract the two and convert to nanoseconds.
*/
- if (intel->gen >= 6)
- query->Base.Result += 80 * (results[1] - results[0]);
- else
- query->Base.Result += 1000 * ((results[1] >> 32) - (results[0] >> 32));
+ query->Base.Result = brw_raw_timestamp_delta(brw, results[0], results[1]);
+ query->Base.Result = brw_timebase_scale(brw, query->Base.Result);
break;
case GL_TIMESTAMP:
/* The query BO contains a single timestamp value in results[0]. */
- if (intel->gen >= 6) {
- /* Our timer is a clock that increments every 80ns (regardless of
- * other clock scaling in the system). The timestamp register we can
- * read for glGetTimestamp() masks out the top 32 bits, so we do that
- * here too to let the two counters be compared against each other.
- *
- * If we just multiplied that 32 bits of data by 80, it would roll
- * over at a non-power-of-two, so an application couldn't use
- * GL_QUERY_COUNTER_BITS to handle rollover correctly. Instead, we
- * report 36 bits and truncate at that (rolling over 5 times as often
- * as the HW counter), and when the 32-bit counter rolls over, it
- * happens to also be at a rollover in the reported value from near
- * (1<<36) to 0.
- *
- * The low 32 bits rolls over in ~343 seconds. Our 36-bit result
- * rolls over every ~69 seconds.
- */
- query->Base.Result = 80 * (results[0] & 0xffffffff);
- query->Base.Result &= (1ull << 36) - 1;
- } else {
- query->Base.Result = 1000 * (results[0] >> 32);
- }
+ query->Base.Result = brw_timebase_scale(brw, results[0]);
+
+ /* Ensure the scaled timestamp overflows according to
+ * GL_QUERY_COUNTER_BITS
+ */
+ query->Base.Result &= (1ull << ctx->Const.QueryCounterBits.Timestamp) - 1;
break;
case GL_SAMPLES_PASSED_ARB:
}
break;
- 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.
- */
- assert(!"Unreachable");
- 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);
+ brw_bo_unmap(query->bo);
/* Now that we've processed the data stored in the query's buffer object,
* we can release it.
*/
- drm_intel_bo_unreference(query->bo);
+ brw_bo_unreference(query->bo);
query->bo = NULL;
}
{
struct brw_query_object *query = (struct brw_query_object *)q;
- drm_intel_bo_unreference(query->bo);
+ brw_bo_unreference(query->bo);
free(query);
}
/**
- * Driver hook for glBeginQuery().
+ * Gen4-5 driver hook for glBeginQuery().
*
* Initializes driver structures and emits any GPU commands required to begin
* recording data for the query.
brw_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;
+ assert(brw->gen < 6);
+
switch (query->Base.Target) {
case GL_TIME_ELAPSED_EXT:
/* 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_bo_unreference(query->bo);
+ query->bo = brw_bo_alloc(brw->bufmgr, "timer query", 4096, 4096);
+ brw_write_timestamp(brw, query->bo, 0);
break;
case GL_ANY_SAMPLES_PASSED:
* 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);
+ brw_bo_unreference(query->bo);
query->bo = NULL;
query->last_index = -1;
* avoid them when necessary. They're required for occlusion queries,
* so turn them on now.
*/
- intel->stats_wm++;
- 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;
- 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;
+ brw->stats_wm++;
+ brw->ctx.NewDriverState |= BRW_NEW_STATS_WM;
break;
default:
- assert(!"Unrecognized query target in brw_begin_query()");
- break;
+ unreachable("Unrecognized query target in brw_begin_query()");
}
}
/**
- * Driver hook for glEndQuery().
+ * Gen4-5 driver hook for glEndQuery().
*
* Emits GPU commands to record a final query value, ending any data capturing.
* However, the final result isn't necessarily available until the GPU processes
brw_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;
+ assert(brw->gen < 6);
+
switch (query->Base.Target) {
case GL_TIME_ELAPSED_EXT:
/* Write the final timestamp. */
- write_timestamp(intel, query->bo, 1);
+ brw_write_timestamp(brw, query->bo, 1);
break;
case GL_ANY_SAMPLES_PASSED:
brw->query.obj = NULL;
- intel->stats_wm--;
- 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;
-
- /* And set brw->query.obj to NULL so that this query won't try to wait
- * for any rendering to complete.
- */
- query->bo = NULL;
- 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;
-
- /* And set brw->query.obj to NULL so that this query won't try to wait
- * for any rendering to complete.
- */
- query->bo = NULL;
+ brw->stats_wm--;
+ brw->ctx.NewDriverState |= BRW_NEW_STATS_WM;
break;
default:
- assert(!"Unrecognized query target in brw_end_query()");
- break;
+ unreachable("Unrecognized query target in brw_end_query()");
}
}
/**
- * The WaitQuery() driver hook.
+ * The Gen4-5 WaitQuery() driver hook.
*
* Wait for a query result to become available and return it. This is the
* backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
{
struct brw_query_object *query = (struct brw_query_object *)q;
+ assert(brw_context(ctx)->gen < 6);
+
brw_queryobj_get_results(ctx, query);
query->Base.Ready = true;
}
/**
- * The CheckQuery() driver hook.
+ * The Gen4-5 CheckQuery() driver hook.
*
* Checks whether a query result is ready yet. If not, flushes.
* This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
*/
static void brw_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;
+ assert(brw->gen < 6);
+
/* 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);
+ if (query->bo && brw_batch_references(&brw->batch, query->bo))
+ intel_batchbuffer_flush(brw);
- if (query->bo == NULL || !drm_intel_bo_busy(query->bo)) {
+ if (query->bo == NULL || !brw_bo_busy(query->bo)) {
brw_queryobj_get_results(ctx, query);
query->Base.Ready = true;
}
static void
ensure_bo_has_space(struct gl_context *ctx, struct brw_query_object *query)
{
- struct intel_context *intel = intel_context(ctx);
+ struct brw_context *brw = brw_context(ctx);
+
+ assert(brw->gen < 6);
if (!query->bo || query->last_index * 2 + 1 >= 4096 / sizeof(uint64_t)) {
brw_queryobj_get_results(ctx, query);
}
- query->bo = drm_intel_bo_alloc(intel->bufmgr, "query", 4096, 1);
+ query->bo = brw_bo_alloc(brw->bufmgr, "query", 4096, 1);
query->last_index = 0;
}
}
* produces the final expected value.
*
* In a world with hardware contexts, PS_DEPTH_COUNT is saved and restored
- * as part of the context state, so this is unnecessary. We could simply
- * read two values and subtract them. However, it's safe to continue using
- * the old approach.
+ * as part of the context state, so this is unnecessary, and skipped.
*/
void
brw_emit_query_begin(struct brw_context *brw)
{
- struct intel_context *intel = &brw->intel;
- struct gl_context *ctx = &intel->ctx;
+ struct gl_context *ctx = &brw->ctx;
struct brw_query_object *query = brw->query.obj;
+ if (brw->hw_ctx)
+ return;
+
/* Skip if we're not doing any queries, or we've already recorded the
* initial query value for this batchbuffer.
*/
ensure_bo_has_space(ctx, query);
- write_depth_count(intel, query->bo, query->last_index * 2);
+ brw_write_depth_count(brw, query->bo, query->last_index * 2);
brw->query.begin_emitted = true;
}
/**
- * Called at batchbuffer flush to get an ending PS_DEPTH_COUNT.
+ * Called at batchbuffer flush to get an ending PS_DEPTH_COUNT
+ * (for non-hardware context platforms).
*
* See the explanation in brw_emit_query_begin().
*/
void
brw_emit_query_end(struct brw_context *brw)
{
- struct intel_context *intel = &brw->intel;
struct brw_query_object *query = brw->query.obj;
+ if (brw->hw_ctx)
+ return;
+
if (!brw->query.begin_emitted)
return;
- write_depth_count(intel, query->bo, query->last_index * 2 + 1);
+ brw_write_depth_count(brw, query->bo, query->last_index * 2 + 1);
brw->query.begin_emitted = false;
query->last_index++;
* current GPU time. This is unlike GL_TIME_ELAPSED, which measures the
* time while the query is active.
*/
-static void
+void
brw_query_counter(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;
assert(q->Target == GL_TIMESTAMP);
- drm_intel_bo_unreference(query->bo);
- query->bo = drm_intel_bo_alloc(intel->bufmgr, "timestamp query", 4096, 4096);
- write_timestamp(intel, query->bo, 0);
+ brw_bo_unreference(query->bo);
+ query->bo = brw_bo_alloc(brw->bufmgr, "timestamp query", 4096, 4096);
+ brw_write_timestamp(brw, query->bo, 0);
+
+ query->flushed = false;
}
/**
static uint64_t
brw_get_timestamp(struct gl_context *ctx)
{
- struct intel_context *intel = intel_context(ctx);
+ struct brw_context *brw = brw_context(ctx);
uint64_t result = 0;
- drm_intel_reg_read(intel->bufmgr, TIMESTAMP, &result);
+ switch (brw->screen->hw_has_timestamp) {
+ case 3: /* New kernel, always full 36bit accuracy */
+ brw_reg_read(brw->bufmgr, TIMESTAMP | 1, &result);
+ break;
+ case 2: /* 64bit kernel, result is left-shifted by 32bits, losing 4bits */
+ brw_reg_read(brw->bufmgr, TIMESTAMP, &result);
+ result = result >> 32;
+ break;
+ case 1: /* 32bit kernel, result is 36bit wide but may be inaccurate! */
+ brw_reg_read(brw->bufmgr, TIMESTAMP, &result);
+ break;
+ }
- /* See logic in brw_queryobj_get_results() */
- result = result >> 32;
- result *= 80;
- result &= (1ull << 36) - 1;
+ /* Scale to nanosecond units */
+ result = brw_timebase_scale(brw, result);
+ /* Ensure the scaled timestamp overflows according to
+ * GL_QUERY_COUNTER_BITS. Technically this isn't required if
+ * querying GL_TIMESTAMP via glGetInteger but it seems best to keep
+ * QueryObject and GetInteger timestamps consistent.
+ */
+ result &= (1ull << ctx->Const.QueryCounterBits.Timestamp) - 1;
return result;
}
-void brw_init_queryobj_functions(struct dd_function_table *functions)
+/**
+ * Is this type of query written by PIPE_CONTROL?
+ */
+bool
+brw_is_query_pipelined(struct brw_query_object *query)
+{
+ switch (query->Base.Target) {
+ case GL_TIMESTAMP:
+ case GL_TIME_ELAPSED:
+ case GL_ANY_SAMPLES_PASSED:
+ case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
+ case GL_SAMPLES_PASSED_ARB:
+ return true;
+
+ case GL_PRIMITIVES_GENERATED:
+ case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
+ case GL_TRANSFORM_FEEDBACK_STREAM_OVERFLOW_ARB:
+ case GL_TRANSFORM_FEEDBACK_OVERFLOW_ARB:
+ 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:
+ return false;
+
+ default:
+ unreachable("Unrecognized query target in is_query_pipelined()");
+ }
+}
+
+/* Initialize query object functions used on all generations. */
+void brw_init_common_queryobj_functions(struct dd_function_table *functions)
{
functions->NewQueryObject = brw_new_query_object;
functions->DeleteQuery = brw_delete_query;
+ functions->GetTimestamp = brw_get_timestamp;
+}
+
+/* Initialize Gen4/5-specific query object functions. */
+void gen4_init_queryobj_functions(struct dd_function_table *functions)
+{
functions->BeginQuery = brw_begin_query;
functions->EndQuery = brw_end_query;
- functions->QueryCounter = brw_query_counter;
functions->CheckQuery = brw_check_query;
functions->WaitQuery = brw_wait_query;
- functions->GetTimestamp = brw_get_timestamp;
+ functions->QueryCounter = brw_query_counter;
}