cuda/vulkan : bicubic interpolation (#17022)

* vulkan : implement upscale with bicubic interpolation

* cuda : implement upscale with bicubic interpolation

* tests : add ggml_interpolate with GGML_SCALE_MODE_BICUBIC to backend tests

* adapt OpenCL backend to not support the OP in that case so tests don't fail

* print scale mode & flags in test-backend-ops

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

@@ -620,7 +620,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_add_id_f32;
 
     vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32;
-    vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32;
+    vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32, pipeline_upscale_bicubic_f32;
     vk_pipeline pipeline_scale_f32;
     vk_pipeline pipeline_sqr_f32;
     vk_pipeline pipeline_sqrt_f32;
@@ -3702,6 +3702,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_upscale_nearest_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_NEAREST}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_upscale_bilinear_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_BILINEAR}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_upscale_bicubic_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_BICUBIC}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
@@ -8200,6 +8201,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
                     return ctx->device->pipeline_upscale_nearest_f32;
                 case GGML_SCALE_MODE_BILINEAR:
                     return ctx->device->pipeline_upscale_bilinear_f32;
+                case GGML_SCALE_MODE_BICUBIC:
+                    return ctx->device->pipeline_upscale_bicubic_f32;
                 default:
                     return nullptr;
             }

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

@@ -20,6 +20,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 // from ggml.h: enum ggml_scale_mode, enum ggml_scale_flag
 #define NEAREST  0
 #define BILINEAR 1
+#define BICUBIC  2
 
 layout (constant_id = 0) const uint scale_mode = 0;
 
@@ -61,6 +62,39 @@ float interpolate_bilinear(uint i10, uint i11, uint i12, uint i13) {
     return fetch_bilinear(c0, c1, d, i12, i13);
 }
 
+// Bicubic interpolation with alpha = -0.75
+// https://en.wikipedia.org/wiki/Bicubic_interpolation#Bicubic_convolution_algorithm
+const vec4 bcoeffs1 = vec4( 1.25, -2.25,  0.0, 1.0);
+const vec4 bcoeffs2 = vec4(-0.75,  3.75, -6.0, 3.0);
+vec4 powers(float x) { return vec4(x*x*x, x*x, x, 1); }
+
+float bicubic(float p0, float p1, float p2, float p3, float x) {
+    return p0 * dot(bcoeffs2, powers(x + 1)) +
+           p1 * dot(bcoeffs1, powers(x    )) +
+           p2 * dot(bcoeffs1, powers(1 - x)) +
+           p3 * dot(bcoeffs2, powers(2 - x));
+}
+
+#define FETCH(a,b) data_a[base + clamp(i.x+(a), 0, res.x) * p.nb00 + clamp(i.y+(b), 0, res.y) * p.nb01]
+
+float interpolate_bicubic(uint i10, uint i11, uint i12, uint i13) {
+    const ivec2 res = ivec2(p.ne00 - 1, p.ne01 - 1);
+
+    const vec2 coord = (vec2(i10, i11) + p.pixel_offset) / vec2(p.sf0, p.sf1) - p.pixel_offset;
+    const vec2 d = fract(coord);
+    const ivec2 i = ivec2(floor(coord));
+
+    const uint i02 = uint(i12 / p.sf2);
+    const uint i03 = uint(i13 / p.sf3);
+    const uint base = p.a_offset + i03 * p.nb03 + i02 * p.nb02;
+
+    return bicubic(
+        bicubic(FETCH(-1,-1), FETCH(0,-1), FETCH(1,-1), FETCH(2,-1), d.x),
+        bicubic(FETCH(-1, 0), FETCH(0, 0), FETCH(1, 0), FETCH(2, 0), d.x),
+        bicubic(FETCH(-1, 1), FETCH(0, 1), FETCH(1, 1), FETCH(2, 1), d.x),
+        bicubic(FETCH(-1, 2), FETCH(0, 2), FETCH(1, 2), FETCH(2, 2), d.x), d.y);
+}
+
 void main() {
     const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
 
@@ -81,6 +115,9 @@ void main() {
         case BILINEAR:
             result = interpolate_bilinear(i10, i11, i12, i13);
             break;
+        case BICUBIC:
+            result = interpolate_bicubic(i10, i11, i12, i13);
+            break;
     }
 
     data_d[p.d_offset + idx] = D_TYPE(result);