ogl_beamforming

Commit: c880d8118d965d5898e6295a19aade47ca05e66c
Parent: cf2955efa08fba1c7f3541621bcb8cf12130cbe9
Author: Randy Palamar
Date:   Tue,  8 Sep 2026 21:25:06 -0700

decode/filter: cleanup over-complicated shared memory loading

This is much more idiomatic and is more likley to allow the
compiler to optimize to using instructions which use the local
tid "register" directly.

Diffstat:
M
shaders/decode.glsl
 | 
11
+----------
M
shaders/filter.glsl
 | 
12
++----------

2 files changed, 3 insertions(+), 20 deletions(-)
diff --git a/shaders/decode.glsl b/shaders/decode.glsl
@@ -28,18 +28,9 @@ void run_decode_large(void)
 	u32 channel     = gl_GlobalInvocationID.y;
 	u32 time_sample = gl_GlobalInvocationID.z;
 
-	const u32 samples_per_thread = TransmitCount / gl_WorkGroupSize.x;
-	const u32 leftover_samples   = TransmitCount % gl_WorkGroupSize.x;
-
-	u32 thread_index_x      = gl_LocalInvocationID.x;
-	u32 samples_this_thread = samples_per_thread + u32(thread_index_x < leftover_samples);
-
 	u32 rf_offset = TransmitCount * ChunkChannelCount * gl_WorkGroupID.z + TransmitCount * channel;
-
-	for (u32 i = 0; i < samples_this_thread; i++) {
-		u32 index = i * gl_WorkGroupSize.x + thread_index_x;
+	for (u32 index = gl_LocalInvocationID.x; index < TransmitCount; index += gl_WorkGroupSize.x)
 		rf[gl_LocalInvocationID.y][index] = RF(rf_buffer).x[rf_offset + index];
-	}
 
 	barrier();
 
diff --git a/shaders/filter.glsl b/shaders/filter.glsl
@@ -66,8 +66,6 @@ void main()
 	uint channel    = gl_GlobalInvocationID.y;
 	uint transmit   = gl_GlobalInvocationID.z;
 
-	uint thread_index = gl_LocalInvocationIndex;
-	uint thread_count = gl_WorkGroupSize.x * gl_WorkGroupSize.y * gl_WorkGroupSize.z;
 	/////////////////////////
 	// NOTE: sample caching
 	{
@@ -86,14 +84,8 @@ void main()
 		input_address += InputDataKindByteSize * (DecimationRate * gl_WorkGroupID.x * gl_WorkGroupSize.x);
 		input_address -= InputDataKindByteSize * (FilterLength - 1);
 
-		uint total_samples       = rf.length();
-		uint samples_per_thread  = total_samples / thread_count;
-		uint leftover_count      = total_samples % thread_count;
-		uint samples_this_thread = samples_per_thread + uint(thread_index < leftover_count);
-
 		const SAMPLE_TYPE scale = SAMPLE_TYPE(bool(ComplexFilter) ? 1 : sqrt(2.0f));
-		for (uint i = 0; i < samples_this_thread; i++) {
-			uint index = thread_count * i + thread_index;
+		for (u32 index = gl_LocalInvocationIndex; index < rf.length(); index += gl_WorkGroupSize.x) {
 			SAMPLE_TYPE s = SAMPLE_TYPE(0);
 			if (!offset_wraps || index >= FilterLength - 1) {
 				s = SAMPLE_TYPE(Input(input_address).x[index]);
@@ -109,7 +101,7 @@ void main()
 	Filter f = Filter(HeapBase + FilterCoefficients);
 	if (out_sample < SampleCount / DecimationRate) {
 		RESULT_TYPE result = RESULT_TYPE(0);
-		u32 offset = DecimationRate * thread_index;
+		u32 offset = DecimationRate * gl_LocalInvocationIndex;
 		for (u32 j = 0; j < FilterLength; j++)
 			result += apply_filter(rf[offset + j], f.values[j]);