Commit: dd2cf92dec5f3c9c588f53e77a4d8613198654c2
Parent: 8214c03d64c9c5348385142591d03670257e0f39
Author: Randy Palamar
Date: Wed, 26 Feb 2025 14:54:52 -0700
core: remove oversized raw_data_ssbo
I don't think it was ever being used correctly and now it is
pointless due to us calling glFinish() before trying to move onto
another frame. If at some point the data is not being loaded by
the worker thread then the fences could be reintroduced but I
don't really see the point.
Diffstat:
2 files changed, 18 insertions(+), 49 deletions(-)
diff --git a/beamformer.c b/beamformer.c
@@ -158,12 +158,11 @@ alloc_shader_storage(BeamformerCtx *ctx, Arena a)
break;
}
- size rf_raw_size = rf_raw_dim.x * rf_raw_dim.y * sizeof(i16);
- size full_rf_buf_size = ARRAY_COUNT(cs->raw_data_fences) * rf_raw_size;
+ size rf_raw_size = rf_raw_dim.x * rf_raw_dim.y * sizeof(i16);
glDeleteBuffers(1, &cs->raw_data_ssbo);
glCreateBuffers(1, &cs->raw_data_ssbo);
- glNamedBufferStorage(cs->raw_data_ssbo, full_rf_buf_size, 0, storage_flags);
+ glNamedBufferStorage(cs->raw_data_ssbo, rf_raw_size, 0, storage_flags);
LABEL_GL_OBJECT(GL_BUFFER, cs->raw_data_ssbo, s8("Raw_RF_SSBO"));
size rf_decoded_size = decoded_data_size(cs);
@@ -184,11 +183,11 @@ alloc_shader_storage(BeamformerCtx *ctx, Arena a)
case GL_VENDOR_ARM:
case GL_VENDOR_INTEL:
cs->raw_data_arena.beg = glMapNamedBufferRange(cs->raw_data_ssbo, 0,
- full_rf_buf_size, map_flags);
- cs->raw_data_arena.end = cs->raw_data_arena.beg + full_rf_buf_size;
+ rf_raw_size, map_flags);
+ cs->raw_data_arena.end = cs->raw_data_arena.beg + rf_raw_size;
break;
case GL_VENDOR_NVIDIA:
- cs->raw_data_arena = ctx->platform.alloc_arena(cs->raw_data_arena, full_rf_buf_size);
+ cs->raw_data_arena = ctx->platform.alloc_arena(cs->raw_data_arena, rf_raw_size);
ctx->cuda_lib.register_cuda_buffers(cs->rf_data_ssbos, ARRAY_COUNT(cs->rf_data_ssbos),
cs->raw_data_ssbo);
ctx->cuda_lib.init_cuda_configuration(bp->rf_raw_dim.E, bp->dec_data_dim.E,
@@ -357,12 +356,9 @@ compute_cursor_finished(struct compute_cursor *cursor)
}
static void
-do_compute_shader(BeamformerCtx *ctx, Arena arena, BeamformFrame *frame, u32 raw_data_index,
- enum compute_shaders shader)
+do_compute_shader(BeamformerCtx *ctx, Arena arena, BeamformFrame *frame, enum compute_shaders shader)
{
ComputeShaderCtx *csctx = &ctx->csctx;
- uv2 rf_raw_dim = ctx->params->raw.rf_raw_dim;
- size rf_raw_size = rf_raw_dim.x * rf_raw_dim.y * sizeof(i16);
glUseProgram(csctx->programs[shader]);
@@ -371,20 +367,16 @@ do_compute_shader(BeamformerCtx *ctx, Arena arena, BeamformFrame *frame, u32 raw
switch (shader) {
case CS_HADAMARD:
- glBindBufferRange(GL_SHADER_STORAGE_BUFFER, 1, csctx->raw_data_ssbo,
- raw_data_index * rf_raw_size, rf_raw_size);
-
+ glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 1, csctx->raw_data_ssbo);
glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 2, csctx->rf_data_ssbos[output_ssbo_idx]);
glBindImageTexture(0, csctx->hadamard_texture, 0, GL_FALSE, 0, GL_READ_ONLY, GL_R8I);
glDispatchCompute(ORONE(csctx->dec_data_dim.x / 32),
ORONE(csctx->dec_data_dim.y / 32),
ORONE(csctx->dec_data_dim.z));
- csctx->raw_data_fences[raw_data_index] = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
csctx->last_output_ssbo_index = !csctx->last_output_ssbo_index;
break;
case CS_CUDA_DECODE:
- ctx->cuda_lib.cuda_decode(raw_data_index * rf_raw_size, output_ssbo_idx, 0);
- csctx->raw_data_fences[raw_data_index] = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
+ ctx->cuda_lib.cuda_decode(0, output_ssbo_idx, 0);
csctx->last_output_ssbo_index = !csctx->last_output_ssbo_index;
break;
case CS_CUDA_HILBERT:
@@ -578,36 +570,14 @@ DEBUG_EXPORT BEAMFORMER_COMPLETE_COMPUTE_FN(beamformer_complete_compute)
#undef X
} break;
case BW_LOAD_RF_DATA: {
- u32 raw_index = cs->raw_data_index;
- if (cs->raw_data_fences[raw_index]) {
- GLsync fence = cs->raw_data_fences[raw_index];
- i32 status = glClientWaitSync(fence, 0, 0);
- if (status != GL_ALREADY_SIGNALED) {
- ctx->platform.write_file(ctx->platform.error_file_handle,
- s8("stall while loading RF data\n"));
- u64 timeout = ctx->gl.max_server_wait_time;
- for (;;) {
- status = glClientWaitSync(fence, 0, timeout);
- if (status == GL_CONDITION_SATISFIED ||
- status == GL_ALREADY_SIGNALED)
- {
- break;
- }
- }
- }
- glDeleteSync(cs->raw_data_fences[raw_index]);
- cs->raw_data_fences[raw_index] = 0;
- }
-
if (!uv2_equal(cs->rf_raw_dim, bp->rf_raw_dim) ||
!uv4_equal(cs->dec_data_dim, bp->dec_data_dim))
{
alloc_shader_storage(ctx, arena);
}
- uv2 rf_raw_dim = cs->rf_raw_dim;
- size rf_raw_size = rf_raw_dim.x * rf_raw_dim.y * sizeof(i16);
- void *rf_data_buf = cs->raw_data_arena.beg + raw_index * rf_raw_size;
+ size rf_raw_size = cs->rf_raw_dim.x * cs->rf_raw_dim.y * sizeof(i16);
+ void *rf_data_buf = cs->raw_data_arena.beg;
size rlen = ctx->platform.read_file(work->file_handle, rf_data_buf, rf_raw_size);
if (rlen != rf_raw_size) {
@@ -626,8 +596,7 @@ DEBUG_EXPORT BEAMFORMER_COMPLETE_COMPUTE_FN(beamformer_complete_compute)
case GL_VENDOR_INTEL:
break;
case GL_VENDOR_NVIDIA:
- glNamedBufferSubData(cs->raw_data_ssbo, raw_index * rlen,
- rlen, rf_data_buf);
+ glNamedBufferSubData(cs->raw_data_ssbo, 0, rlen, rf_data_buf);
}
}
ctx->ready_for_rf = 1;
@@ -660,7 +629,7 @@ DEBUG_EXPORT BEAMFORMER_COMPLETE_COMPUTE_FN(beamformer_complete_compute)
for (u32 i = 0; i < stage_count; i++) {
frame->timer_active[stages[i]] = 1;
glBeginQuery(GL_TIME_ELAPSED, frame->timer_ids[stages[i]]);
- do_compute_shader(ctx, arena, frame, cs->raw_data_index, stages[i]);
+ do_compute_shader(ctx, arena, frame, stages[i]);
glEndQuery(GL_TIME_ELAPSED);
}
/* NOTE(rnp): block until work completes so that we can record timings */
diff --git a/beamformer.h b/beamformer.h
@@ -108,8 +108,6 @@ typedef struct {
v2 last_mouse;
} BeamformerInput;
-#define MAX_FRAMES_IN_FLIGHT 3
-
#define INIT_CUDA_CONFIGURATION_FN(name) void name(u32 *input_dims, u32 *decoded_dims, u16 *channel_mapping)
typedef INIT_CUDA_CONFIGURATION_FN(init_cuda_configuration_fn);
INIT_CUDA_CONFIGURATION_FN(init_cuda_configuration_stub) {}
@@ -196,10 +194,12 @@ 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[MAX_FRAMES_IN_FLIGHT];
- Arena raw_data_arena;
- u32 raw_data_ssbo;
- u32 raw_data_index;
+ /* NOTE: The raw data ssbo is not mapped on NVIDIA because their drivers _will_ store
+ * the buffer in the system memory. This doesn't happen for other vendors and
+ * mapping the buffer is preferred. In either case incoming data can be written to
+ * the arena. An additional BufferSubData is needed on NVIDIA to upload the data. */
+ Arena raw_data_arena;
+ u32 raw_data_ssbo;
/* NOTE: Decoded data is only relevant in the context of a single frame. We use two
* buffers so that they can be swapped when chaining multiple compute stages */