X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fmesa%2Fdrivers%2Fdri%2Fi965%2Fbrw_queryobj.c;h=b27a685d052ef9169365e9c4679e01c13bdc6d28;hb=e5339fe4a47c242693962c9f90bbab8b74935cba;hp=985dbedddfca54a7f21d4121242cac1d7662c234;hpb=647fc0c50bc9832c336b2b7e4329abec31df9dec;p=mesa.git diff --git a/src/mesa/drivers/dri/i965/brw_queryobj.c b/src/mesa/drivers/dri/i965/brw_queryobj.c index 985dbedddfc..b27a685d052 100644 --- a/src/mesa/drivers/dri/i965/brw_queryobj.c +++ b/src/mesa/drivers/dri/i965/brw_queryobj.c @@ -35,82 +35,88 @@ * appropriately synced with the stage of the pipeline for our extensions' * needs. */ -#include "main/imports.h" +#include "util/imports.h" +#include "main/queryobj.h" #include "brw_context.h" #include "brw_defines.h" #include "brw_state.h" #include "intel_batchbuffer.h" -#include "intel_reg.h" + +/* 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.) + */ +uint64_t +brw_raw_timestamp_delta(struct brw_context *brw, uint64_t time0, uint64_t time1) +{ + 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; + } + } +} /** * 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, struct brw_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(); - } + const struct gen_device_info *devinfo = &brw->screen->devinfo; - 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 (devinfo->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 (devinfo->gen == 9 && devinfo->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) { - assert(intel->gen < 6); - - 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(); + const struct gen_device_info *devinfo = &brw->screen->devinfo; + uint32_t flags = PIPE_CONTROL_WRITE_DEPTH_COUNT | PIPE_CONTROL_DEPTH_STALL; + + if (devinfo->gen == 9 && devinfo->gt == 4) + flags |= PIPE_CONTROL_CS_STALL; + + if (devinfo->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." + */ + brw_emit_pipe_control_flush(brw, PIPE_CONTROL_DEPTH_STALL); + } + + brw_emit_pipe_control_write(brw, flags, + query_bo, idx * sizeof(uint64_t), 0); } /** @@ -120,12 +126,13 @@ 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); + UNUSED const struct gen_device_info *devinfo = &brw->screen->devinfo; int i; uint64_t *results; - assert(intel->gen < 6); + assert(devinfo->gen < 6); if (query->bo == NULL) return; @@ -134,28 +141,33 @@ 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 (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. */ - 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 = gen_device_info_timebase_scale(devinfo, query->Base.Result); break; case GL_TIMESTAMP: /* The query BO contains a single timestamp value in results[0]. */ - query->Base.Result = 1000 * (results[0] >> 32); + query->Base.Result = gen_device_info_timebase_scale(devinfo, 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: @@ -188,15 +200,14 @@ brw_queryobj_get_results(struct gl_context *ctx, 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; } @@ -228,8 +239,8 @@ brw_delete_query(struct gl_context *ctx, struct gl_query_object *q) { struct brw_query_object *query = (struct brw_query_object *)q; - drm_intel_bo_unreference(query->bo); - free(query); + brw_bo_unreference(query->bo); + _mesa_delete_query(ctx, q); } /** @@ -242,10 +253,10 @@ 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; + UNUSED const struct gen_device_info *devinfo = &brw->screen->devinfo; - assert(intel->gen < 6); + assert(devinfo->gen < 6); switch (query->Base.Target) { case GL_TIME_ELAPSED_EXT: @@ -268,9 +279,10 @@ brw_begin_query(struct gl_context *ctx, struct gl_query_object *q) * 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, BRW_MEMZONE_OTHER); + brw_write_timestamp(brw, query->bo, 0); break; case GL_ANY_SAMPLES_PASSED: @@ -283,7 +295,7 @@ brw_begin_query(struct gl_context *ctx, struct gl_query_object *q) * 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; @@ -293,13 +305,12 @@ 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++; - brw->state.dirty.brw |= BRW_NEW_STATS_WM; + 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()"); } } @@ -315,15 +326,15 @@ 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; + UNUSED const struct gen_device_info *devinfo = &brw->screen->devinfo; - assert(intel->gen < 6); + assert(devinfo->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: @@ -352,13 +363,12 @@ brw_end_query(struct gl_context *ctx, struct gl_query_object *q) brw->query.obj = NULL; - intel->stats_wm--; - brw->state.dirty.brw |= BRW_NEW_STATS_WM; + 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()"); } } @@ -371,8 +381,10 @@ brw_end_query(struct gl_context *ctx, struct gl_query_object *q) static void brw_wait_query(struct gl_context *ctx, struct gl_query_object *q) { struct brw_query_object *query = (struct brw_query_object *)q; + UNUSED const struct gen_device_info *devinfo = + &brw_context(ctx)->screen->devinfo; - assert(intel_context(ctx)->gen < 6); + assert(devinfo->gen < 6); brw_queryobj_get_results(ctx, query); query->Base.Ready = true; @@ -386,10 +398,11 @@ static void brw_wait_query(struct gl_context *ctx, struct gl_query_object *q) */ 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; + UNUSED const struct gen_device_info *devinfo = &brw->screen->devinfo; - assert(intel->gen < 6); + assert(devinfo->gen < 6); /* From the GL_ARB_occlusion_query spec: * @@ -398,10 +411,10 @@ 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 && 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; } @@ -416,9 +429,10 @@ 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); + UNUSED const struct gen_device_info *devinfo = &brw->screen->devinfo; - assert(intel->gen < 6); + assert(devinfo->gen < 6); if (!query->bo || query->last_index * 2 + 1 >= 4096 / sizeof(uint64_t)) { @@ -430,7 +444,7 @@ 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); + query->bo = brw_bo_alloc(brw->bufmgr, "query", 4096, BRW_MEMZONE_OTHER); query->last_index = 0; } } @@ -458,13 +472,9 @@ ensure_bo_has_space(struct gl_context *ctx, struct brw_query_object *query) 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 (intel->hw_ctx) - return; - /* Skip if we're not doing any queries, or we've already recorded the * initial query value for this batchbuffer. */ @@ -473,7 +483,7 @@ 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; } @@ -487,16 +497,12 @@ brw_emit_query_begin(struct brw_context *brw) void brw_emit_query_end(struct brw_context *brw) { - struct intel_context *intel = &brw->intel; struct brw_query_object *query = brw->query.obj; - if (intel->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++; @@ -509,17 +515,20 @@ 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); + brw_bo_unreference(query->bo); + query->bo = + brw_bo_alloc(brw->bufmgr, "timestamp query", 4096, BRW_MEMZONE_OTHER); + brw_write_timestamp(brw, query->bo, 0); + + query->flushed = false; } /** @@ -530,25 +539,76 @@ 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); + const struct gen_device_info *devinfo = &brw->screen->devinfo; 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 = gen_device_info_timebase_scale(devinfo, 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; } +/** + * 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->QueryCounter = brw_query_counter; functions->GetTimestamp = brw_get_timestamp; } @@ -559,4 +619,5 @@ void gen4_init_queryobj_functions(struct dd_function_table *functions) functions->EndQuery = brw_end_query; functions->CheckQuery = brw_check_query; functions->WaitQuery = brw_wait_query; + functions->QueryCounter = brw_query_counter; }