vulkan: add ABS op

Signed-off-by: Giuseppe Scrivano <[email protected]>

Author: giuseppe

docs/ops.md

@@ -14,7 +14,7 @@ Legend:
 
 | Operation | BLAS | CANN | CPU | CUDA | Metal | OpenCL | SYCL | Vulkan | zDNN |
 |-----------|------|------|------|------|------|------|------|------|------|
-|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ✅ | ❌ |
 |                              ACC | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
 |                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -659,6 +659,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_sigmoid[2];
     vk_pipeline pipeline_hardsigmoid[2];
     vk_pipeline pipeline_hardswish[2];
+    vk_pipeline pipeline_abs[2];
 
     vk_pipeline pipeline_geglu[2];
     vk_pipeline pipeline_reglu[2];
@@ -3735,6 +3736,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     CREATE_UNARY(sigmoid)
     CREATE_UNARY(hardsigmoid)
     CREATE_UNARY(hardswish)
+    CREATE_UNARY(abs)
 #undef CREATE_UNARY
 
 #define CREATE_UNARY_RTE(name)  \
@@ -8334,6 +8336,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
                 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];
+            case GGML_UNARY_OP_ABS:
+                return ctx->device->pipeline_abs[dst->type == GGML_TYPE_F16];
             default:
                 break;
         }
@@ -11286,6 +11290,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_UNARY_OP_SIGMOID:
         case GGML_UNARY_OP_HARDSIGMOID:
         case GGML_UNARY_OP_HARDSWISH:
+        case GGML_UNARY_OP_ABS:
             break;
         default:
             return false;
@@ -11617,6 +11622,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_UNARY_OP_SIGMOID:
         case GGML_UNARY_OP_HARDSIGMOID:
         case GGML_UNARY_OP_HARDSWISH:
+        case GGML_UNARY_OP_ABS:
             ggml_vk_unary(ctx, compute_ctx, src0, node);
             break;
         default:
@@ -11888,6 +11894,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
         case GGML_UNARY_OP_SIGMOID:
         case GGML_UNARY_OP_HARDSIGMOID:
         case GGML_UNARY_OP_HARDSWISH:
+        case GGML_UNARY_OP_ABS:
             buf = tensor->buffer;
             break;
         default:
@@ -13383,6 +13390,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 case GGML_UNARY_OP_SIGMOID:
                 case GGML_UNARY_OP_HARDSIGMOID:
                 case GGML_UNARY_OP_HARDSWISH:
+                case GGML_UNARY_OP_ABS:
                     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) &&

ggml/src/ggml-vulkan/vulkan-shaders/abs.comp

@@ -0,0 +1,21 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+#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;
+    }
+
+    data_d[i] = D_TYPE(abs(float(data_a[i])));
+}

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -837,6 +837,8 @@ void process_shaders() {
     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"}});
+    string_to_spv("abs_f16",        "abs.comp",         {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("abs_f32",        "abs.comp",         {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
 
     for (auto rte : {false, true}) {
         std::string suffix = rte ? "_rte" : "";