ogl_beamforming

Commit: af12ff47508553f07ce7c1f9785892c5248e99ad
Parent: 16c104c39dd46f04a190170a2755fd612685dafc
Author: Randy Palamar
Date:   Thu, 15 Aug 2024 10:51:31 -0600

make sure there is a set of timers for each frame in flight

Diffstat:
M
beamformer.c
 | 
11
+++++------
M
beamformer.h
 | 
8
+++++---
M
main.c
 | 
2
+-

3 files changed, 11 insertions(+), 10 deletions(-)
diff --git a/beamformer.c b/beamformer.c
@@ -96,7 +96,7 @@ do_compute_shader(BeamformerCtx *ctx, enum compute_shaders shader)
 	uv2  rf_raw_dim         = ctx->params->raw.rf_raw_dim;
 	size rf_raw_size        = rf_raw_dim.x * rf_raw_dim.y * sizeof(i16);
 
-	glBeginQuery(GL_TIME_ELAPSED, csctx->timer_ids[shader]);
+	glBeginQuery(GL_TIME_ELAPSED, csctx->timer_ids[csctx->timer_index][shader]);
 
 	glUseProgram(csctx->programs[shader]);
 
@@ -509,14 +509,13 @@ check_compute_timers(ComputeShaderCtx *cs)
 	if (!cs->timer_fences[last_idx])
 		return;
 
-	i32 timer_status, _unused;
-	glGetSynciv(cs->timer_fences[last_idx], GL_SYNC_STATUS, 4, &_unused, &timer_status);
-	if (timer_status != GL_SIGNALED)
+	i32 status = glClientWaitSync(cs->timer_fences[last_idx], 0, 0);
+	if (status == GL_TIMEOUT_EXPIRED || status == GL_WAIT_FAILED)
 		return;
 
-	for (u32 i = 0; i < ARRAY_COUNT(cs->timer_ids); i++) {
+	for (u32 i = 0; i < ARRAY_COUNT(cs->last_frame_time); i++) {
 		u64 ns = 0;
-		glGetQueryObjectui64v(cs->timer_ids[i], GL_QUERY_RESULT, &ns);
+		glGetQueryObjectui64v(cs->timer_ids[last_idx][i], GL_QUERY_RESULT, &ns);
 		cs->last_frame_time[i] = (f32)ns / 1e9;
 	}
 }
diff --git a/beamformer.h b/beamformer.h
@@ -94,12 +94,14 @@ typedef struct {
 	#error Unsupported Platform!
 #endif
 
+#define MAX_FRAMES_IN_FLIGHT 3
+
 typedef struct {
 	u32 programs[CS_LAST];
 
 	u32    timer_index;
-	u32    timer_ids[CS_LAST];
-	GLsync timer_fences[3];
+	u32    timer_ids[MAX_FRAMES_IN_FLIGHT][CS_LAST];
+	GLsync timer_fences[MAX_FRAMES_IN_FLIGHT];
 	f32    last_frame_time[CS_LAST];
 
 	/* NOTE: the raw_data_ssbo is allocated at 3x the required size to allow for tiled
@@ -108,7 +110,7 @@ typedef struct {
 	 * for Intel or AMD and mapping the buffer is preferred. In either case incoming data can
 	 * be written to the arena at the appropriate offset for the current raw_data_index. An
 	 * additional BufferSubData is needed on NVIDIA to upload the data. */
-	GLsync raw_data_fences[3];
+	GLsync raw_data_fences[MAX_FRAMES_IN_FLIGHT];
 	Arena  raw_data_arena;
 	u32    raw_data_ssbo;
 	u32    raw_data_index;
diff --git a/main.c b/main.c
@@ -207,7 +207,7 @@ main(void)
 	glCreateBuffers(1, &ctx.csctx.shared_ubo);
 	glNamedBufferStorage(ctx.csctx.shared_ubo, sizeof(BeamformerParameters), 0, GL_DYNAMIC_STORAGE_BIT);
 
-	glGenQueries(CS_LAST, ctx.csctx.timer_ids);
+	glGenQueries(ARRAY_COUNT(ctx.csctx.timer_fences) * CS_LAST, (u32 *)ctx.csctx.timer_ids);
 
 	/* NOTE: do not DO_COMPUTE on first frame */
 	reload_shaders(&ctx, temp_memory);