ogl_beamforming

Commit: e0bd75572cdba7626c8dcf6d2fb87f8aa77660f7
Parent: ad11111303575137d78ca4118bb1601bcc11b69d
Author: Randy Palamar
Date:   Tue,  3 Mar 2026 10:13:45 -0700

decode/filter: use f16 as intermediate exchange type

This gives ~20% speed up for the decoding.

Diffstat:
M
beamformer_core.c
 | 
2
+-
M
shaders/decode.glsl
 | 
6
++++++
M
shaders/filter.glsl
 | 
13
+++++++------

3 files changed, 14 insertions(+), 7 deletions(-)
diff --git a/beamformer_core.c b/beamformer_core.c
@@ -365,7 +365,7 @@ plan_compute_pipeline(BeamformerComputePlan *cp, BeamformerParameterBlock *pb)
 				db->data_kind = data_kind;
 				if (!first) {
 					if (data_kind == BeamformerDataKind_Int16) {
-						db->data_kind = BeamformerDataKind_Int16Complex;
+						db->data_kind = BeamformerDataKind_Float16Complex;
 					} else {
 						db->data_kind = BeamformerDataKind_Float32Complex;
 					}
diff --git a/shaders/decode.glsl b/shaders/decode.glsl
@@ -14,6 +14,12 @@
 #elif DataKind == DataKind_Float32Complex
   #define INPUT_DATA_TYPE  vec2
   #define SAMPLE_DATA_TYPE vec2
+#elif DataKind == DataKind_Float16Complex
+  #define INPUT_DATA_TYPE  f16vec2
+  #define SAMPLE_DATA_TYPE vec2
+#elif DataKind == DataKind_Float16
+  #define INPUT_DATA_TYPE  float16_t
+  #define SAMPLE_DATA_TYPE float
 #elif DataKind == DataKind_Int16Complex
   #define INPUT_DATA_TYPE  i16vec2
   #define SAMPLE_DATA_TYPE vec2
diff --git a/shaders/filter.glsl b/shaders/filter.glsl
@@ -1,18 +1,19 @@
 /* See LICENSE for license details. */
 /* TODO(rnp): bug: this won't filter RF data correctly */
-#define SAMPLE_TYPE vec2
+#define SAMPLE_TYPE f32vec2
 #if DataKind == DataKind_Float32
-  #define DATA_TYPE           vec2
+  #define DATA_TYPE           f32vec2
   #define RESULT_TYPE_CAST(v) (v)
   #define SAMPLE_TYPE_CAST(v) (v)
 #else
-  #define DATA_TYPE           uint
-  #define SAMPLE_TYPE_CAST(v) unpackSnorm2x16(v)
+  #define DATA_TYPE           i16vec2
+  #define SAMPLE_TYPE_CAST(v) (v)
   #if OutputFloats
-    #define OUT_DATA_TYPE       vec2
+    #define OUT_DATA_TYPE       f32vec2
     #define RESULT_TYPE_CAST(v) (clamp((v), -1.0, 1.0) * 32767.0f)
   #else
-    #define RESULT_TYPE_CAST(v) packSnorm2x16(v)
+    #define OUT_DATA_TYPE       f16vec2
+    #define RESULT_TYPE_CAST(v) f16vec2(v)
   #endif
 #endif