ogl_beamforming

Commit: 3939648169851306c32bf37a52f73ac8567c85fb
Parent: ba9b99fed19c9da46620c15c04639080dae659a6
Author: Randy Palamar
Date:   Mon,  3 Nov 2025 06:10:20 -0700

shaders/decode: add a compile time flag for LDS

Diffstat:
M
beamformer.c
 | 
8
+++++---
M
beamformer.meta
 | 
2
+-
M
generated/beamformer.meta.c
 | 
6
++++--
M
shaders/decode.glsl
 | 
4
++--

4 files changed, 12 insertions(+), 8 deletions(-)
diff --git a/beamformer.c b/beamformer.c
@@ -505,8 +505,6 @@ plan_compute_pipeline(BeamformerComputePlan *cp, BeamformerParameterBlock *pb)
 				}
 			}
 
-			if (run_cuda_hilbert) sd->bake.flags |= BeamformerShaderDecodeFlags_DilateOutput;
-
 			BeamformerShaderKind *last_shader = cp->pipeline.shaders + slot - 1;
 			assert(first || ((*last_shader == BeamformerShaderKind_Demodulate ||
 			                  *last_shader == BeamformerShaderKind_Filter)));
@@ -527,9 +525,13 @@ plan_compute_pipeline(BeamformerComputePlan *cp, BeamformerParameterBlock *pb)
 				db->output_transmit_stride *= decimation_rate;
 			}
 
+			if (run_cuda_hilbert)        sd->bake.flags |= BeamformerShaderDecodeFlags_DilateOutput;
+			if (db->transmit_count > 32) sd->bake.flags |= BeamformerShaderDecodeFlags_UseSharedMemory;
+
 			db->transmits_processed = db->transmit_count >= 32 ? 2 : 1;
 
-			b32 use_16z  = db->transmit_count <= 32 || db->transmit_count == 80 || db->transmit_count == 96 || db->transmit_count == 160;
+			b32 use_16z  = db->transmit_count <= 32 || db->transmit_count == 80 ||
+			               db->transmit_count == 96 || db->transmit_count == 160;
 			sd->layout.x = 4;
 			sd->layout.y = 1;
 			sd->layout.z = use_16z? 16 : 32;
diff --git a/beamformer.meta b/beamformer.meta
@@ -59,7 +59,7 @@
 	{
 		@Enumeration(DataKind)
 		@Enumeration(DecodeMode)
-		@Flags([DilateOutput])
+		@Flags([DilateOutput UseSharedMemory])
 
 		@Bake
 		{
diff --git a/generated/beamformer.meta.c b/generated/beamformer.meta.c
@@ -53,7 +53,8 @@ typedef enum {
 } BeamformerInterpolationMode;
 
 typedef enum {
-	BeamformerShaderDecodeFlags_DilateOutput = (1 << 0),
+	BeamformerShaderDecodeFlags_DilateOutput    = (1 << 0),
+	BeamformerShaderDecodeFlags_UseSharedMemory = (1 << 1),
 } BeamformerShaderDecodeFlags;
 
 typedef enum {
@@ -240,6 +241,7 @@ read_only global s8 beamformer_shader_global_header_strings[] = {
 read_only global s8 *beamformer_shader_flag_strings[] = {
 	(s8 []){
 		s8_comp("DilateOutput"),
+		s8_comp("UseSharedMemory"),
 	},
 	(s8 []){
 		s8_comp("ComplexFilter"),
@@ -259,7 +261,7 @@ read_only global s8 *beamformer_shader_flag_strings[] = {
 };
 
 read_only global u8 beamformer_shader_flag_strings_count[] = {
-	1,
+	2,
 	3,
 	5,
 	0,
diff --git a/shaders/decode.glsl b/shaders/decode.glsl
@@ -90,7 +90,7 @@ void main()
 					result[i] = sample_rf_data(rf_offset + transmit + i);
 		}break;
 		case DecodeMode_Hadamard:{
-			#if TransmitCount > 32
+			#if UseSharedMemory
 			{
 				uint thread_count = gl_WorkGroupSize.x * gl_WorkGroupSize.y * gl_WorkGroupSize.z;
 				uint thread_index = gl_LocalInvocationIndex;
@@ -115,7 +115,7 @@ void main()
 					result[i] = SAMPLE_DATA_TYPE(0);
 
 				for (int j = 0; j < TransmitCount; j++) {
-					#if TransmitCount > 32
+					#if UseSharedMemory
 					SAMPLE_DATA_TYPE s = SAMPLE_TYPE_CAST(rf[gl_LocalInvocationID.x * TransmitCount + j]);
 					#else
 					SAMPLE_DATA_TYPE s = sample_rf_data(rf_offset + j);