From: Shin-myoung-serp <redacted>
Date: Wed, 3 Sep 2025 18:22:55 +0000 (+0900)
Subject: ggml vulkan: add hardsigmoid and hardswish operations (#15762)
X-Git-Tag: upstream/0.0.6527~156
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=0014fb4add3a7d000c14b763538045c6170f57d9;p=pkg%2Fggml%2Fsources%2Fllama.cpp

ggml vulkan: add hardsigmoid and hardswish operations (#15762)
---

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 7189fc1c..ea34bd26 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -529,6 +529,8 @@ struct vk_device_struct {
     vk_pipeline pipeline_relu[2];
     vk_pipeline pipeline_tanh[2];
     vk_pipeline pipeline_sigmoid[2];
+    vk_pipeline pipeline_hardsigmoid[2];
+    vk_pipeline pipeline_hardswish[2];
 
     vk_pipeline pipeline_geglu[2];
     vk_pipeline pipeline_reglu[2];
@@ -3261,6 +3263,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
     CREATE_UNARY(relu)
     CREATE_UNARY(tanh)
     CREATE_UNARY(sigmoid)
+    CREATE_UNARY(hardsigmoid)
+    CREATE_UNARY(hardswish)
 #undef CREATE_UNARY
 
 #define CREATE_GLU(name)  \
@@ -7533,6 +7537,10 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
                 return ctx->device->pipeline_tanh[dst->type == GGML_TYPE_F16];
             case GGML_UNARY_OP_SIGMOID:
                 return ctx->device->pipeline_sigmoid[dst->type == GGML_TYPE_F16];
+            case GGML_UNARY_OP_HARDSIGMOID:
+                return ctx->device->pipeline_hardsigmoid[dst->type == GGML_TYPE_F16];
+            case GGML_UNARY_OP_HARDSWISH:
+                return ctx->device->pipeline_hardswish[dst->type == GGML_TYPE_F16];
             default:
                 break;
         }
@@ -10201,6 +10209,8 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_UNARY_OP_RELU:
         case GGML_UNARY_OP_TANH:
         case GGML_UNARY_OP_SIGMOID:
+        case GGML_UNARY_OP_HARDSIGMOID:
+        case GGML_UNARY_OP_HARDSWISH:
             break;
         default:
             return false;
@@ -10571,6 +10581,8 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_UNARY_OP_RELU:
         case GGML_UNARY_OP_TANH:
         case GGML_UNARY_OP_SIGMOID:
+        case GGML_UNARY_OP_HARDSIGMOID:
+        case GGML_UNARY_OP_HARDSWISH:
             ggml_vk_unary(ctx, compute_ctx, src0, node, dryrun);
             break;
         default:
@@ -10813,6 +10825,8 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
         case GGML_UNARY_OP_RELU:
         case GGML_UNARY_OP_TANH:
         case GGML_UNARY_OP_SIGMOID:
+        case GGML_UNARY_OP_HARDSIGMOID:
+        case GGML_UNARY_OP_HARDSWISH:
             buf = tensor->buffer;
             break;
         default:
@@ -11764,6 +11778,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 case GGML_UNARY_OP_RELU:
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_SIGMOID:
+                case GGML_UNARY_OP_HARDSIGMOID:
+                case GGML_UNARY_OP_HARDSWISH:
                     return ggml_is_contiguous(op->src[0]) &&
                            (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
                            (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) &&
@@ -12580,6 +12596,12 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         case GGML_UNARY_OP_SIGMOID:
             tensor_clone = ggml_sigmoid(ggml_ctx, src_clone[0]);
             break;
+        case GGML_UNARY_OP_HARDSIGMOID:
+            tensor_clone = ggml_hardsigmoid(ggml_ctx, src_clone[0]);
+            break;
+        case GGML_UNARY_OP_HARDSWISH:
+            tensor_clone = ggml_hardswish(ggml_ctx, src_clone[0]);
+            break;
         default:
             std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
             GGML_ABORT("fatal error");
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/hardsigmoid.comp b/ggml/src/ggml-vulkan/vulkan-shaders/hardsigmoid.comp
new file mode 100644
index 00000000..1da252cc
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/hardsigmoid.comp
@@ -0,0 +1,22 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float x = float(data_a[i]);
+    data_d[i] = D_TYPE(min(1.0f, max(0.0f, (x + 3.0f) / 6.0f)));
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/hardswish.comp b/ggml/src/ggml-vulkan/vulkan-shaders/hardswish.comp
new file mode 100644
index 00000000..3afc5882
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/hardswish.comp
@@ -0,0 +1,22 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float x = float(data_a[i]);
+    data_d[i] = D_TYPE(x * min(1.0f, max(0.0f, (x + 3.0f) / 6.0f)));
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index 1263a70e..613498d0 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -657,6 +657,10 @@ void process_shaders() {
     string_to_spv("tanh_f32",       "tanh.comp",        {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
     string_to_spv("sigmoid_f16",    "sigmoid.comp",     {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("sigmoid_f32",    "sigmoid.comp",     {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    string_to_spv("hardsigmoid_f16","hardsigmoid.comp", {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("hardsigmoid_f32","hardsigmoid.comp", {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    string_to_spv("hardswish_f16",  "hardswish.comp",   {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("hardswish_f32",  "hardswish.comp",   {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
 
     for (auto rte : {false, true}) {
         std::string suffix = rte ? "_rte" : "";