X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fmesa%2Fdrivers%2Fdri%2Fi965%2Fbrw_queryobj.c;h=a8e5ababba47c8f509e2d3467e27322221523d2d;hb=65a5af6dd0550e93a5250582147689a1ddef9ffa;hp=a1a1353dee7605c354be61842400fe2b601375ac;hpb=295d6f8e8f03192320aa8d4ed767427dd06071a5;p=mesa.git diff --git a/src/mesa/drivers/dri/i965/brw_queryobj.c b/src/mesa/drivers/dri/i965/brw_queryobj.c index a1a1353dee7..a8e5ababba4 100644 --- a/src/mesa/drivers/dri/i965/brw_queryobj.c +++ b/src/mesa/drivers/dri/i965/brw_queryobj.c @@ -25,237 +25,500 @@ * */ -/** @file support for ARB_query_object +/** @file brw_queryobj.c * - * ARB_query_object is implemented by using the PIPE_CONTROL command to stall - * execution on the completion of previous depth tests, and write the - * current PS_DEPTH_COUNT to a buffer object. + * Support for query objects (GL_ARB_occlusion_query, GL_ARB_timer_query, + * GL_EXT_transform_feedback, and friends). * - * We use before and after counts when drawing during a query so that - * we don't pick up other clients' query data in ours. To reduce overhead, - * a single BO is used to record the query data for all active queries at - * once. This also gives us a simple bound on how much batchbuffer space is - * required for handling queries, so that we can be sure that we won't - * have to emit a batchbuffer without getting the ending PS_DEPTH_COUNT. + * The hardware provides a PIPE_CONTROL command that can report the number of + * fragments that passed the depth test, or the hardware timer. They are + * appropriately synced with the stage of the pipeline for our extensions' + * needs. */ -#include "main/simple_list.h" #include "main/imports.h" #include "brw_context.h" +#include "brw_defines.h" +#include "brw_state.h" #include "intel_batchbuffer.h" #include "intel_reg.h" -/** Waits on the query object's BO and totals the results for this query */ +/** + * Emit PIPE_CONTROLs to write the current GPU timestamp into a buffer. + */ +void +brw_write_timestamp(struct brw_context *brw, drm_intel_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); + } + + brw_emit_pipe_control_write(brw, PIPE_CONTROL_WRITE_TIMESTAMP, + query_bo, idx * sizeof(uint64_t), 0, 0); +} + +/** + * Emit PIPE_CONTROLs to write the PS_DEPTH_COUNT register into a buffer. + */ +void +brw_write_depth_count(struct brw_context *brw, drm_intel_bo *query_bo, int idx) +{ + brw_emit_pipe_control_write(brw, + PIPE_CONTROL_WRITE_DEPTH_COUNT | + PIPE_CONTROL_DEPTH_STALL, + query_bo, idx * sizeof(uint64_t), + 0, 0); +} + +/** + * Wait on the query object's BO and calculate the final result. + */ static void -brw_queryobj_get_results(struct brw_query_object *query) +brw_queryobj_get_results(struct gl_context *ctx, + struct brw_query_object *query) { + struct brw_context *brw = brw_context(ctx); + int i; uint64_t *results; + assert(brw->gen < 6); + if (query->bo == NULL) return; - /* Map and count the pixels from the current query BO */ - dri_bo_map(query->bo, GL_FALSE); + /* 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(brw->batch.bo, query->bo)) + intel_batchbuffer_flush(brw); + + if (unlikely(brw->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); results = query->bo->virtual; - for (i = query->first_index; i <= query->last_index; i++) { - query->Base.Result += results[i * 2 + 1] - results[i * 2]; + 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. + */ + query->Base.Result += 1000 * ((results[1] >> 32) - (results[0] >> 32)); + break; + + case GL_TIMESTAMP: + /* The query BO contains a single timestamp value in results[0]. */ + query->Base.Result = 1000 * (results[0] >> 32); + break; + + case GL_SAMPLES_PASSED_ARB: + /* Loop over pairs of values from the BO, which are the PS_DEPTH_COUNT + * value at the start and end of the batchbuffer. Subtract them to + * get the number of fragments which passed the depth test in each + * individual batch, and add those differences up to get the number + * of fragments for the entire query. + * + * Note that query->Base.Result may already be non-zero. We may have + * run out of space in the query's BO and allocated a new one. If so, + * this function was already called to accumulate the results so far. + */ + for (i = 0; i < query->last_index; i++) { + query->Base.Result += results[i * 2 + 1] - results[i * 2]; + } + break; + + case GL_ANY_SAMPLES_PASSED: + case GL_ANY_SAMPLES_PASSED_CONSERVATIVE: + /* If the starting and ending PS_DEPTH_COUNT from any of the batches + * differ, then some fragments passed the depth test. + */ + for (i = 0; i < query->last_index; i++) { + if (results[i * 2 + 1] != results[i * 2]) { + query->Base.Result = GL_TRUE; + break; + } + } + break; + + default: + unreachable("Unrecognized query target in brw_queryobj_get_results()"); } - dri_bo_unmap(query->bo); + drm_intel_bo_unmap(query->bo); - dri_bo_unreference(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); query->bo = NULL; } +/** + * The NewQueryObject() driver hook. + * + * Allocates and initializes a new query object. + */ static struct gl_query_object * -brw_new_query_object(GLcontext *ctx, GLuint id) +brw_new_query_object(struct gl_context *ctx, GLuint id) { struct brw_query_object *query; - query = _mesa_calloc(sizeof(struct brw_query_object)); + query = calloc(1, sizeof(struct brw_query_object)); query->Base.Id = id; query->Base.Result = 0; - query->Base.Active = GL_FALSE; - query->Base.Ready = GL_TRUE; + query->Base.Active = false; + query->Base.Ready = true; return &query->Base; } +/** + * The DeleteQuery() driver hook. + */ static void -brw_delete_query(GLcontext *ctx, struct gl_query_object *q) +brw_delete_query(struct gl_context *ctx, struct gl_query_object *q) { struct brw_query_object *query = (struct brw_query_object *)q; - dri_bo_unreference(query->bo); - _mesa_free(query); + drm_intel_bo_unreference(query->bo); + free(query); } +/** + * Gen4-5 driver hook for glBeginQuery(). + * + * Initializes driver structures and emits any GPU commands required to begin + * recording data for the query. + */ static void -brw_begin_query(GLcontext *ctx, struct gl_query_object *q) +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; - /* Reset our driver's tracking of query state. */ - dri_bo_unreference(query->bo); - query->bo = NULL; - query->first_index = -1; - query->last_index = -1; - - insert_at_head(&brw->query.active_head, query); - intel->stats_wm++; + 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 + * we measure the full time between BeginQuery and EndQuery. There's + * some debate about whether this is the right thing to do. Our decision + * is based on the following text from the ARB_timer_query extension: + * + * "(5) Should the extension measure total time elapsed between the full + * completion of the BeginQuery and EndQuery commands, or just time + * spent in the graphics library? + * + * RESOLVED: This extension will measure the total time elapsed + * between the full completion of these commands. Future extensions + * may implement a query to determine time elapsed at different stages + * of the graphics pipeline." + * + * We write a starting timestamp now (at index 0). At EndQuery() time, + * we'll write a second timestamp (at index 1), and subtract the two to + * 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(brw->bufmgr, "timer query", 4096, 4096); + brw_write_timestamp(brw, query->bo, 0); + break; + + case GL_ANY_SAMPLES_PASSED: + case GL_ANY_SAMPLES_PASSED_CONSERVATIVE: + case GL_SAMPLES_PASSED_ARB: + /* For occlusion queries, we delay taking an initial sample until the + * first drawing occurs in this batch. See the reasoning in the comments + * for brw_emit_query_begin() below. + * + * 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 = NULL; + query->last_index = -1; + + brw->query.obj = query; + + /* Depth statistics on Gen4 require strange workarounds, so we try to + * avoid them when necessary. They're required for occlusion queries, + * so turn them on now. + */ + brw->stats_wm++; + brw->ctx.NewDriverState |= BRW_NEW_STATS_WM; + break; + + default: + unreachable("Unrecognized query target in brw_begin_query()"); + } } /** - * Begin the ARB_occlusion_query query on a query object. + * 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 + * those commands. brw_queryobj_get_results() processes the captured data to + * produce the final result. */ static void -brw_end_query(GLcontext *ctx, struct gl_query_object *q) +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; - /* Flush the batchbuffer in case it has writes to our query BO. - * Have later queries write to a new query BO so that further rendering - * doesn't delay the collection of our results. - */ - if (query->bo) { + assert(brw->gen < 6); + + switch (query->Base.Target) { + case GL_TIME_ELAPSED_EXT: + /* Write the final timestamp. */ + brw_write_timestamp(brw, query->bo, 1); + break; + + case GL_ANY_SAMPLES_PASSED: + case GL_ANY_SAMPLES_PASSED_CONSERVATIVE: + case GL_SAMPLES_PASSED_ARB: + + /* No query->bo means that EndQuery was called after BeginQuery with no + * intervening drawing. Rather than doing nothing at all here in this + * case, we emit the query_begin and query_end state to the + * hardware. This is to guarantee that waiting on the result of this + * empty state will cause all previous queries to complete at all, as + * required by the specification: + * + * It must always be true that if any query object + * returns a result available of TRUE, all queries of the + * same type issued prior to that query must also return + * TRUE. [Open GL 4.3 (Core Profile) Section 4.2.1] + */ + if (!query->bo) { + brw_emit_query_begin(brw); + } + + assert(query->bo); + brw_emit_query_end(brw); - intel_batchbuffer_flush(intel->batch); - dri_bo_unreference(brw->query.bo); - brw->query.bo = NULL; - } + brw->query.obj = NULL; - remove_from_list(query); + brw->stats_wm--; + brw->ctx.NewDriverState |= BRW_NEW_STATS_WM; + break; - intel->stats_wm--; + default: + unreachable("Unrecognized query target in brw_end_query()"); + } } -static void brw_wait_query(GLcontext *ctx, struct gl_query_object *q) +/** + * 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. + */ +static void brw_wait_query(struct gl_context *ctx, struct gl_query_object *q) { struct brw_query_object *query = (struct brw_query_object *)q; - brw_queryobj_get_results(query); - query->Base.Ready = GL_TRUE; + assert(brw_context(ctx)->gen < 6); + + brw_queryobj_get_results(ctx, query); + query->Base.Ready = true; } -static void brw_check_query(GLcontext *ctx, struct gl_query_object *q) +/** + * 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) { - /* XXX: Need to expose dri_bo_is_idle from bufmgr. */ -#if 0 + struct brw_context *brw = brw_context(ctx); struct brw_query_object *query = (struct brw_query_object *)q; - if (dri_bo_is_idle(query->bo)) { - brw_queryobj_get_results(query); - query->Base.Ready = GL_TRUE; + assert(brw->gen < 6); + + /* From the GL_ARB_occlusion_query spec: + * + * "Instead of allowing for an infinite loop, performing a + * QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is + * 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(brw->batch.bo, query->bo)) + intel_batchbuffer_flush(brw); + + if (query->bo == NULL || !drm_intel_bo_busy(query->bo)) { + brw_queryobj_get_results(ctx, query); + query->Base.Ready = true; } -#else - brw_wait_query(ctx, q); -#endif } -/** Called to set up the query BO and account for its aperture space */ -void -brw_prepare_query_begin(struct brw_context *brw) +/** + * Ensure there query's BO has enough space to store a new pair of values. + * + * If not, gather the existing BO's results and create a new buffer of the + * same size. + */ +static void +ensure_bo_has_space(struct gl_context *ctx, struct brw_query_object *query) { - struct intel_context *intel = &brw->intel; - dri_bo *aper_array[] = { - intel->batch->buf, - brw->query.bo, - }; - - /* Skip if we're not doing any queries. */ - if (is_empty_list(&brw->query.active_head)) - return; + struct brw_context *brw = brw_context(ctx); - /* Get a new query BO if we're going to need it. */ - if (brw->query.bo == NULL || - brw->query.index * 2 + 1 >= 4096 / sizeof(uint64_t)) { - dri_bo_unreference(brw->query.bo); - brw->query.bo = NULL; + assert(brw->gen < 6); - brw->query.bo = dri_bo_alloc(intel->bufmgr, "query", 4096, 1); - brw->query.index = 0; - } + if (!query->bo || query->last_index * 2 + 1 >= 4096 / sizeof(uint64_t)) { - if (dri_bufmgr_check_aperture_space(aper_array, ARRAY_SIZE(aper_array))) - intel_batchbuffer_flush(intel->batch); + if (query->bo != NULL) { + /* The old query BO did not have enough space, so we allocated a new + * one. Gather the results so far (adding up the differences) and + * release the old BO. + */ + brw_queryobj_get_results(ctx, query); + } + + query->bo = drm_intel_bo_alloc(brw->bufmgr, "query", 4096, 1); + query->last_index = 0; + } } -/** Called just before primitive drawing to get a beginning PS_DEPTH_COUNT. */ +/** + * Record the PS_DEPTH_COUNT value (for occlusion queries) just before + * primitive drawing. + * + * In a pre-hardware context world, the single PS_DEPTH_COUNT register is + * shared among all applications using the GPU. However, our query value + * needs to only include fragments generated by our application/GL context. + * + * To accommodate this, we record PS_DEPTH_COUNT at the start and end of + * each batchbuffer (technically, the first primitive drawn and flush time). + * Subtracting each pair of values calculates the change in PS_DEPTH_COUNT + * caused by a batchbuffer. Since there is no preemption inside batches, + * this is guaranteed to only measure the effects of our current application. + * + * Adding each of these differences (in case drawing is done over many batches) + * 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, and skipped. + */ void brw_emit_query_begin(struct brw_context *brw) { - struct intel_context *intel = &brw->intel; - struct brw_query_object *query; + struct gl_context *ctx = &brw->ctx; + struct brw_query_object *query = brw->query.obj; - /* Skip if we're not doing any queries, or we've emitted the start. */ - if (brw->query.active || is_empty_list(&brw->query.active_head)) + if (brw->hw_ctx) return; - BEGIN_BATCH(4, IGNORE_CLIPRECTS); - OUT_BATCH(_3DSTATE_PIPE_CONTROL | - 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. + /* Skip if we're not doing any queries, or we've already recorded the + * initial query value for this batchbuffer. */ - OUT_RELOC(brw->query.bo, - I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION, - PIPE_CONTROL_GLOBAL_GTT_WRITE | - ((brw->query.index * 2) * sizeof(uint64_t))); - OUT_BATCH(0); - OUT_BATCH(0); - ADVANCE_BATCH(); - - foreach(query, &brw->query.active_head) { - if (query->bo != brw->query.bo) { - if (query->bo != NULL) - brw_queryobj_get_results(query); - dri_bo_reference(brw->query.bo); - query->bo = brw->query.bo; - query->first_index = brw->query.index; - } - query->last_index = brw->query.index; - } - brw->query.active = GL_TRUE; + if (!query || brw->query.begin_emitted) + return; + + ensure_bo_has_space(ctx, query); + + 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.active) + if (!brw->query.begin_emitted) return; - BEGIN_BATCH(4, IGNORE_CLIPRECTS); - OUT_BATCH(_3DSTATE_PIPE_CONTROL | - PIPE_CONTROL_DEPTH_STALL | - PIPE_CONTROL_WRITE_DEPTH_COUNT); - OUT_RELOC(brw->query.bo, - I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION, - PIPE_CONTROL_GLOBAL_GTT_WRITE | - ((brw->query.index * 2 + 1) * sizeof(uint64_t))); - OUT_BATCH(0); - OUT_BATCH(0); - ADVANCE_BATCH(); - - brw->query.active = GL_FALSE; - brw->query.index++; + brw_write_depth_count(brw, query->bo, query->last_index * 2 + 1); + + brw->query.begin_emitted = false; + query->last_index++; } -void brw_init_queryobj_functions(struct dd_function_table *functions) +/** + * Driver hook for glQueryCounter(). + * + * This handles GL_TIMESTAMP queries, which perform a pipelined read of the + * current GPU time. This is unlike GL_TIME_ELAPSED, which measures the + * time while the query is active. + */ +static void +brw_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; + + assert(q->Target == GL_TIMESTAMP); + + drm_intel_bo_unreference(query->bo); + query->bo = drm_intel_bo_alloc(brw->bufmgr, "timestamp query", 4096, 4096); + brw_write_timestamp(brw, query->bo, 0); + + query->flushed = false; +} + +/** + * Read the TIMESTAMP register immediately (in a non-pipelined fashion). + * + * This is used to implement the GetTimestamp() driver hook. + */ +static uint64_t +brw_get_timestamp(struct gl_context *ctx) +{ + struct brw_context *brw = brw_context(ctx); + uint64_t result = 0; + + switch (brw->intelScreen->hw_has_timestamp) { + case 3: /* New kernel, always full 36bit accuracy */ + drm_intel_reg_read(brw->bufmgr, TIMESTAMP | 1, &result); + break; + case 2: /* 64bit kernel, result is left-shifted by 32bits, losing 4bits */ + drm_intel_reg_read(brw->bufmgr, TIMESTAMP, &result); + result = result >> 32; + break; + case 1: /* 32bit kernel, result is 36bit wide but may be inaccurate! */ + drm_intel_reg_read(brw->bufmgr, TIMESTAMP, &result); + break; + } + + /* See logic in brw_queryobj_get_results() */ + result *= 80; + result &= (1ull << 36) - 1; + return result; +} + +/* 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->QueryCounter = brw_query_counter; + 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->CheckQuery = brw_check_query;