X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fmesa%2Fdrivers%2Fdri%2Fi965%2Fbrw_queryobj.c;h=bd3f5738eba6fee5673fa0c1f124a737701a72e3;hb=27e273578f05521baac08d7de915c95312e3a595;hp=d6b012c392e2811ae347aafaa76094b5e5113162;hpb=6a7ca4ef2cd3f39d3b5e77051cb3f3175e9e60df;p=mesa.git diff --git a/src/mesa/drivers/dri/i965/brw_queryobj.c b/src/mesa/drivers/dri/i965/brw_queryobj.c index d6b012c392e..bd3f5738eba 100644 --- a/src/mesa/drivers/dri/i965/brw_queryobj.c +++ b/src/mesa/drivers/dri/i965/brw_queryobj.c @@ -41,45 +41,89 @@ #include "brw_defines.h" #include "brw_state.h" #include "intel_batchbuffer.h" -#include "intel_reg.h" + +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.) + */ +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. */ void -brw_write_timestamp(struct brw_context *brw, drm_intel_bo *query_bo, int idx) +brw_write_timestamp(struct brw_context *brw, struct brw_bo *query_bo, int idx) { - if (brw->gen == 6) { + const struct gen_device_info *devinfo = &brw->screen->devinfo; + + if (devinfo->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); + 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. */ void -brw_write_depth_count(struct brw_context *brw, drm_intel_bo *query_bo, int idx) +brw_write_depth_count(struct brw_context *brw, struct brw_bo *query_bo, int idx) { - uint32_t flags; + const struct gen_device_info *devinfo = &brw->screen->devinfo; + uint32_t flags = PIPE_CONTROL_WRITE_DEPTH_COUNT | PIPE_CONTROL_DEPTH_STALL; - flags = (PIPE_CONTROL_WRITE_DEPTH_COUNT | - PIPE_CONTROL_DEPTH_STALL); + if (devinfo->gen == 9 && devinfo->gt == 4) + flags |= PIPE_CONTROL_CS_STALL; - /* Needed to ensure the memory is coherent for the MI_LOAD_REGISTER_MEM - * command when loading the values into the predicate source registers for - * conditional rendering. - */ - if (brw->predicate.supported) - flags |= PIPE_CONTROL_FLUSH_ENABLE; + 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, 0); + brw_emit_pipe_control_write(brw, flags, + query_bo, idx * sizeof(uint64_t), 0); } /** @@ -90,11 +134,12 @@ brw_queryobj_get_results(struct gl_context *ctx, struct brw_query_object *query) { struct brw_context *brw = brw_context(ctx); + const struct gen_device_info *devinfo = &brw->screen->devinfo; int i; uint64_t *results; - assert(brw->gen < 6); + assert(devinfo->gen < 6); if (query->bo == NULL) return; @@ -103,28 +148,33 @@ 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(brw->batch.bo, query->bo)) + if (brw_batch_references(&brw->batch, query->bo)) intel_batchbuffer_flush(brw); if (unlikely(brw->perf_debug)) { - if (drm_intel_bo_busy(query->bo)) { + 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 = brw_timebase_scale(brw, 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 = 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: @@ -159,12 +209,12 @@ brw_queryobj_get_results(struct gl_context *ctx, default: 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; } @@ -196,7 +246,7 @@ 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); + brw_bo_unreference(query->bo); free(query); } @@ -211,8 +261,9 @@ brw_begin_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; + const struct gen_device_info *devinfo = &brw->screen->devinfo; - assert(brw->gen < 6); + assert(devinfo->gen < 6); switch (query->Base.Target) { case GL_TIME_ELAPSED_EXT: @@ -235,8 +286,8 @@ 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(brw->bufmgr, "timer query", 4096, 4096); + brw_bo_unreference(query->bo); + query->bo = brw_bo_alloc(brw->bufmgr, "timer query", 4096, 4096); brw_write_timestamp(brw, query->bo, 0); break; @@ -250,7 +301,7 @@ 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; @@ -282,8 +333,9 @@ brw_end_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; + const struct gen_device_info *devinfo = &brw->screen->devinfo; - assert(brw->gen < 6); + assert(devinfo->gen < 6); switch (query->Base.Target) { case GL_TIME_ELAPSED_EXT: @@ -335,8 +387,9 @@ 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; + const struct gen_device_info *devinfo = &brw_context(ctx)->screen->devinfo; - assert(brw_context(ctx)->gen < 6); + assert(devinfo->gen < 6); brw_queryobj_get_results(ctx, query); query->Base.Ready = true; @@ -352,8 +405,9 @@ static void brw_check_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; + const struct gen_device_info *devinfo = &brw->screen->devinfo; - assert(brw->gen < 6); + assert(devinfo->gen < 6); /* From the GL_ARB_occlusion_query spec: * @@ -362,10 +416,10 @@ 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(brw->batch.bo, query->bo)) + 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; } @@ -381,8 +435,9 @@ static void ensure_bo_has_space(struct gl_context *ctx, struct brw_query_object *query) { struct brw_context *brw = brw_context(ctx); + const struct gen_device_info *devinfo = &brw->screen->devinfo; - assert(brw->gen < 6); + assert(devinfo->gen < 6); if (!query->bo || query->last_index * 2 + 1 >= 4096 / sizeof(uint64_t)) { @@ -394,7 +449,7 @@ 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(brw->bufmgr, "query", 4096, 1); + query->bo = brw_bo_alloc(brw->bufmgr, "query", 4096, 1); query->last_index = 0; } } @@ -471,7 +526,7 @@ 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 brw_context *brw = brw_context(ctx); @@ -479,8 +534,8 @@ brw_query_counter(struct gl_context *ctx, struct gl_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_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; @@ -497,31 +552,72 @@ brw_get_timestamp(struct gl_context *ctx) struct brw_context *brw = brw_context(ctx); uint64_t result = 0; - switch (brw->intelScreen->hw_has_timestamp) { + switch (brw->screen->hw_has_timestamp) { case 3: /* New kernel, always full 36bit accuracy */ - drm_intel_reg_read(brw->bufmgr, TIMESTAMP | 1, &result); + brw_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); + brw_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); + brw_reg_read(brw->bufmgr, TIMESTAMP, &result); break; } - /* See logic in brw_queryobj_get_results() */ - 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; } +/** + * 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; } @@ -532,4 +628,5 @@ 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; }