X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fmesa%2Fdrivers%2Fdri%2Fi965%2Fbrw_queryobj.c;h=6ba2296371a0fd332eadf2cba245493d2cd2307c;hb=55364ab5b7136e09a61d858f1167dee81e17bd9f;hp=de3be83c8d7ee0edb1133c8f0038a9fc309c18e2;hpb=14cec07177f438717cc6fb9252525e16d6b3d8dd;p=mesa.git diff --git a/src/mesa/drivers/dri/i965/brw_queryobj.c b/src/mesa/drivers/dri/i965/brw_queryobj.c index de3be83c8d7..6ba2296371a 100644 --- a/src/mesa/drivers/dri/i965/brw_queryobj.c +++ b/src/mesa/drivers/dri/i965/brw_queryobj.c @@ -41,93 +41,43 @@ #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. */ -static void -write_timestamp(struct intel_context *intel, drm_intel_bo *query_bo, int idx) +void +brw_write_timestamp(struct brw_context *brw, drm_intel_bo *query_bo, int idx) { - 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(); - } - - 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(); + 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, 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, drm_intel_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. - */ - 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(); - } + uint32_t flags = PIPE_CONTROL_WRITE_DEPTH_COUNT | PIPE_CONTROL_DEPTH_STALL; + + 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, 0); } /** @@ -137,11 +87,13 @@ static void 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; @@ -149,10 +101,10 @@ brw_queryobj_get_results(struct gl_context *ctx, * 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 (drm_intel_bo_references(brw->batch.bo, query->bo)) + intel_batchbuffer_flush(brw); - if (unlikely(intel->perf_debug)) { + if (unlikely(brw->perf_debug)) { if (drm_intel_bo_busy(query->bo)) { perf_debug("Stalling on the GPU waiting for a query object.\n"); } @@ -165,36 +117,12 @@ brw_queryobj_get_results(struct gl_context *ctx, /* 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 += 1000 * ((results[1] >> 32) - (results[0] >> 32)); 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 = 1000 * (results[0] >> 32); break; case GL_SAMPLES_PASSED_ARB: @@ -208,7 +136,7 @@ brw_queryobj_get_results(struct gl_context *ctx, * 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++) { + for (i = 0; i < query->last_index; i++) { query->Base.Result += results[i * 2 + 1] - results[i * 2]; } break; @@ -218,7 +146,7 @@ brw_queryobj_get_results(struct gl_context *ctx, /* 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++) { + for (i = 0; i < query->last_index; i++) { if (results[i * 2 + 1] != results[i * 2]) { query->Base.Result = GL_TRUE; break; @@ -226,17 +154,8 @@ brw_queryobj_get_results(struct gl_context *ctx, } 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); @@ -280,7 +199,7 @@ brw_delete_query(struct gl_context *ctx, struct gl_query_object *q) } /** - * 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. @@ -289,9 +208,10 @@ static void 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 @@ -314,8 +234,8 @@ brw_begin_query(struct gl_context *ctx, struct gl_query_object *q) * 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); + 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: @@ -338,33 +258,17 @@ brw_begin_query(struct gl_context *ctx, struct gl_query_object *q) * 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 @@ -375,13 +279,14 @@ static void 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: @@ -410,45 +315,17 @@ brw_end_query(struct gl_context *ctx, struct gl_query_object *q) 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. @@ -457,21 +334,25 @@ static void brw_wait_query(struct gl_context *ctx, struct gl_query_object *q) { 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 @@ -479,8 +360,8 @@ static void brw_check_query(struct gl_context *ctx, struct gl_query_object *q) * 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 && 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); @@ -497,7 +378,9 @@ static void brw_check_query(struct gl_context *ctx, struct gl_query_object *q) 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)) { @@ -509,13 +392,7 @@ ensure_bo_has_space(struct gl_context *ctx, struct brw_query_object *query) brw_queryobj_get_results(ctx, query); } - query->bo = drm_intel_bo_alloc(intel->bufmgr, "query", 4096, 1); - - /* Fill the buffer with zeroes. This is probably superfluous. */ - drm_intel_bo_map(query->bo, true); - memset((char *) query->bo->virtual, 0, 4096); - drm_intel_bo_unmap(query->bo); - + query->bo = drm_intel_bo_alloc(brw->bufmgr, "query", 4096, 1); query->last_index = 0; } } @@ -538,17 +415,17 @@ ensure_bo_has_space(struct gl_context *ctx, struct brw_query_object *query) * 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. */ @@ -557,26 +434,29 @@ brw_emit_query_begin(struct brw_context *brw) 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++; @@ -589,17 +469,19 @@ brw_emit_query_end(struct brw_context *brw) * 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); + query->bo = drm_intel_bo_alloc(brw->bufmgr, "timestamp query", 4096, 4096); + brw_write_timestamp(brw, query->bo, 0); + + query->flushed = false; } /** @@ -610,27 +492,76 @@ brw_query_counter(struct gl_context *ctx, struct gl_query_object *q) 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->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 = result >> 32; result *= 80; result &= (1ull << 36) - 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_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; }