From: Jeff Bolz <redacted>
Date: Mon, 1 Sep 2025 19:01:10 +0000 (-0500)
Subject: vulkan: add missing clamps in new mul_mat_id paths (llama/15702)
X-Git-Tag: v0.9.1~109
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=2d57866e8786583cc6e1162fc4486ee4c63bd856;p=pkg%2Fggml%2Fsources%2Fggml

vulkan: add missing clamps in new mul_mat_id paths (llama/15702)

This is a missing interaction between #15546 and #15652
---

diff --git a/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp b/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp
index 654105a4..69ac38fd 100644
--- a/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp
+++ b/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp
@@ -494,6 +494,9 @@ void main() {
                     sum = coopMatMulAdd(mat_a, mat_b, sum);
                 }
             }
+#if defined(ACC_TYPE_MAX)
+            [[unroll]] for (uint i = 0; i < sum.length(); ++i) { sum[i] = clamp(sum[i], -ACC_TYPE_MAX, ACC_TYPE_MAX); }
+#endif
 
             // Convert from ACC_TYPE to D_TYPE
             coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BNover4, gl_MatrixUseAccumulator> mat_d;
@@ -535,6 +538,9 @@ void main() {
                     sum = coopMatMulAdd(mat_a, mat_b, sum);
                 }
             }
+#if defined(ACC_TYPE_MAX)
+            [[unroll]] for (uint i = 0; i < sum.length(); ++i) { sum[i] = clamp(sum[i], -ACC_TYPE_MAX, ACC_TYPE_MAX); }
+#endif
 
             // Convert from ACC_TYPE to D_TYPE
             coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BNover2, gl_MatrixUseAccumulator> mat_d;