From: Reese Levine <redacted>
Date: Sun, 5 Oct 2025 03:59:31 +0000 (-0700)
Subject: ggml webgpu: actually add softmax, fix rms_norm offset (llama/16400)
X-Git-Tag: upstream/1.8.2~43
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=b8bdf061829942e7843e74e0988ed0e417ae10f1;p=pkg%2Fggml%2Fsources%2Fwhisper.cpp

ggml webgpu: actually add softmax, fix rms_norm offset (llama/16400)

* implement soft_max

* Fix soft_max data race

* Temporary fix, wait on each submit
---

diff --git a/ggml/src/ggml-webgpu/ggml-webgpu.cpp b/ggml/src/ggml-webgpu/ggml-webgpu.cpp
index de68c568..e795ca3f 100644
--- a/ggml/src/ggml-webgpu/ggml-webgpu.cpp
+++ b/ggml/src/ggml-webgpu/ggml-webgpu.cpp
@@ -424,6 +424,7 @@ static void ggml_backend_webgpu_build_and_enqueue(webgpu_context &
         ctx->staged_param_bufs.push_back(params_bufs);
         if (ctx->staged_command_bufs.size() == WEBGPU_COMMAND_SUBMIT_BATCH_SIZE) {
             ggml_backend_webgpu_submit_queue(ctx);
+            ggml_backend_webgpu_wait_on_submission(ctx);
         }
     }
 }
@@ -1060,6 +1061,9 @@ static bool ggml_webgpu_encode_node(webgpu_context ctx, ggml_tensor * node) {
         case GGML_OP_SCALE:
             ggml_webgpu_scale(ctx, src0, node);
             break;
+        case GGML_OP_SOFT_MAX:
+            ggml_webgpu_soft_max(ctx, src0, src1, src2, node);
+            break;
         default:
             return false;
     }
@@ -1806,6 +1810,9 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
         case GGML_OP_SCALE:
             supports_op = op->type == GGML_TYPE_F32;
             break;
+        case GGML_OP_SOFT_MAX:
+            supports_op = op->type == GGML_TYPE_F32;
+            break;
         default:
             break;
     }
@@ -1949,6 +1956,7 @@ static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t
     ggml_webgpu_init_rope_pipeline(ctx);
     ggml_webgpu_init_glu_pipeline(ctx);
     ggml_webgpu_init_scale_pipeline(ctx);
+    ggml_webgpu_init_soft_max_pipeline(ctx);
 
 #ifdef GGML_WEBGPU_DEBUG
     // Initialize debug buffers
diff --git a/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl b/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl
index 4f72bb1c..712b921f 100644
--- a/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl
+++ b/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl
@@ -84,7 +84,7 @@ fn main(@builtin(workgroup_id) wid: vec3<u32>,
     let i2 = i / params.ne1;
     let i1 = i % params.ne1;
     let i_src_row = params.offset_src + i3 * params.stride_src3 + i2 * params.stride_src2 + i1 * params.stride_src1;
-    let i_dst_row = params.offset_src + i3 * params.stride_dst3 + i2 * params.stride_dst2 + i1 * params.stride_dst1;
+    let i_dst_row = params.offset_dst + i3 * params.stride_dst3 + i2 * params.stride_dst2 + i1 * params.stride_dst1;
 
     let elems = (params.ne0 + wg_size - 1) / wg_size;
 
diff --git a/ggml/src/ggml-webgpu/wgsl-shaders/soft_max.tmpl.wgsl b/ggml/src/ggml-webgpu/wgsl-shaders/soft_max.tmpl.wgsl
index 64ab576c..c74dc4cc 100644
--- a/ggml/src/ggml-webgpu/wgsl-shaders/soft_max.tmpl.wgsl
+++ b/ggml/src/ggml-webgpu/wgsl-shaders/soft_max.tmpl.wgsl
@@ -300,6 +300,7 @@ fn main(@builtin(workgroup_id) wid: vec3<u32>,
         workgroupBarrier();
     }
     let row_max = scratch[0];
+    workgroupBarrier();
 
     var sum = 0.0f;
     col = lid.x;