From: cmdr2 <redacted>
Date: Fri, 28 Feb 2025 10:29:55 +0000 (+0530)
Subject: cuda/vulkan: specify fp32-only support for some operations in supports_op (#1129)
X-Git-Tag: upstream/0.0.1802~38
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=ff9052988b76e137bcf92bb335733933ca196ac0;p=pkg%2Fggml%2Fsources%2Fggml

cuda/vulkan: specify fp32-only support for some operations in supports_op (#1129)

* cuda: restrict SILU_BACK to fp32, since fp16 exceeds the desired test threshold

* vulkan: specify fp32-only support for certain ops (that are now tested for fp16 as well)

* f32 sigmoid in vulkan supports op

* Revert "f32 sigmoid in vulkan supports op"

This reverts commit c6f04b3c19bf4504c2776149c6d8cd84e0b48acb.
---

diff --git a/src/ggml-cuda/ggml-cuda.cu b/src/ggml-cuda/ggml-cuda.cu
index d4948057..fe30259f 100644
--- a/src/ggml-cuda/ggml-cuda.cu
+++ b/src/ggml-cuda/ggml-cuda.cu
@@ -3153,7 +3153,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 return false;
             } break;
         case GGML_OP_SILU_BACK:
-            return ggml_is_contiguous(op->src[0]);
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
             break;
         case GGML_OP_NORM:
         case GGML_OP_RMS_NORM:
diff --git a/src/ggml-vulkan/ggml-vulkan.cpp b/src/ggml-vulkan/ggml-vulkan.cpp
index 131ee1ea..e0066c36 100644
--- a/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/src/ggml-vulkan/ggml-vulkan.cpp
@@ -8371,7 +8371,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_RELU:
                 case GGML_UNARY_OP_TANH:
-                    return ggml_is_contiguous(op->src[0]);
+                    return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
                 default:
                     return false;
             }
@@ -8571,17 +8571,18 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_RMS_NORM:
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_ADD:
-        case GGML_OP_ACC:
         case GGML_OP_SUB:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
-        case GGML_OP_CONCAT:
-        case GGML_OP_UPSCALE:
-        case GGML_OP_SCALE:
         case GGML_OP_SQR:
         case GGML_OP_SIN:
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
+            return op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_ACC:
+        case GGML_OP_CONCAT:
+        case GGML_OP_UPSCALE:
+        case GGML_OP_SCALE:
         case GGML_OP_PAD:
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 461ef91c..d5cd7dbf 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -3980,10 +3980,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
 
     test_cases.emplace_back(new test_add1());
     test_cases.emplace_back(new test_scale());
-
-    for (ggml_type type : {GGML_TYPE_F16, GGML_TYPE_F32}) {
-        test_cases.emplace_back(new test_silu_back());
-    }
+    test_cases.emplace_back(new test_silu_back());
 
     for (float eps : {0.0f, 1e-6f, 1e-4f, 1e-1f}) {
         for (bool v : {false, true}) {