[ET-VK][Ops] aten.convolution (Transpose)

Pull Request resolved: #2883

## Summary (cases handled)

We introduce support for the convolution cases covered by ATen-VK's transpose implementation. This is achieved by
- reusing the existing [`conv_transpose2d.glsl`](https://github.com/pytorch/pytorch/blob/09c72eaa3f69f90402c86a30abf4fc621298578c/aten/src/ATen/native/vulkan/glsl/conv_transpose2d.glsl), and
- [moving special weights prepacking from CPU](https://github.com/pytorch/pytorch/blob/09c72eaa3f69f90402c86a30abf4fc621298578c/aten/src/ATen/native/vulkan/ops/Convolution.cpp#L134-L235) to the GPU in `conv_transpose2d_prepack_weights.glsl`.

We also include resizing support for dynamic shapes. Note that only height and width of the input can vary.

## Cases not handled

The implementation is on-par with ATen-VK's Transpose. This means the following cases are missing:
1. **Groups G > 1.**
2. **Batch (input) N > 1.**
3. **Dilation > 1.**
ghstack-source-id: 221721754
@exported-using-ghexport

Differential Revision: [D55667336](https://our.internmc.facebook.com/intern/diff/D55667336/)

Author: jorgep31415

backends/vulkan/runtime/graph/ops/glsl/conv_transpose2d.glsl

@@ -0,0 +1,92 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#version 450 core
+
+#define PRECISION ${PRECISION}
+
+layout(std430) buffer;
+
+layout(set = 0, binding = 0, ${IMAGE_FORMAT[DTYPE]}) uniform PRECISION restrict writeonly ${IMAGE_T[NDIM][DTYPE]} image_out;
+layout(set = 0, binding = 1) uniform PRECISION sampler3D image_in;
+layout(set = 0, binding = 2) uniform PRECISION sampler2D kernel_in;
+layout(set = 0, binding = 3) uniform PRECISION sampler2D bias_in;
+
+layout(set = 0, binding = 4) uniform PRECISION restrict OutExtents {
+  uvec4 data;
+}
+out_extents;
+
+layout(set = 0, binding = 5) uniform PRECISION restrict InExtents {
+  uvec4 data;
+}
+in_extents;
+
+layout(set = 0, binding = 6) uniform PRECISION restrict Params {
+  ivec2 kernel_size;
+  ivec2 stride;
+  ivec2 padding;
+  ivec2 dilation;
+}
+params;
+
+// If fields are separated, SwiftShader cannot identify in_group_size.
+layout(set = 0, binding = 7) uniform PRECISION restrict ExtraParams {
+  ivec2 overlay_region;
+  int in_group_size;
+}
+extra_params;
+
+layout(local_size_x_id = 0, local_size_y_id = 1, local_size_z_id = 2) in;
+
+/*
+ * Computes a 2D transpose convolution. Each shader invocation calculates the
+ * output at a single output location. For details, refer to conv2d.glsl which
+ * uses a similar approach.
+ */
+void main() {
+  const ivec3 pos = ivec3(gl_GlobalInvocationID);
+
+  if (any(greaterThanEqual(pos, out_extents.data.xyz))) {
+    return;
+  }
+
+  ivec2 ipos = pos.xy + params.padding;
+
+  const ivec2 start = max(
+      ivec2(0),
+      ivec2(ceil((vec2(ipos) - params.kernel_size + 1) / vec2(params.stride))));
+  const ivec2 end =
+      min(ivec2(in_extents.data.xy),
+          ivec2(floor(vec2(ipos) / vec2(params.stride))) + 1);
+
+  const int ic = extra_params.in_group_size;
+  const int kx_stride = ic * (params.stride.x - 1);
+
+  int ky_start = extra_params.overlay_region.y - 1 -
+      (ipos.y - params.stride.y * start.y) + pos.z * params.kernel_size.y;
+  int kx_start = (extra_params.overlay_region.x - 1 -
+                  (ipos.x - params.stride.x * start.x)) * ic;
+
+  ${VEC4_T[DTYPE]} sum = texelFetch(bias_in, ivec2(pos.z, 0), 0);
+  for (int y = start.y, ky = ky_start; y < end.y; ++y, ky += params.stride.y) {
+    for (int x = start.x, kx = kx_start; x < end.x; ++x, kx += kx_stride) {
+      for (int z4 = 0; z4 < ic / 4; ++z4, kx += 4) {
+        const ${VEC4_T[DTYPE]} in_texel = texelFetch(image_in, ivec3(x, y, z4), 0);
+        const ivec4 kxs = kx + ivec4(0, 1, 2, 3);
+
+        sum = fma(in_texel.xxxx, texelFetch(kernel_in, ivec2(kxs.x, ky), 0), sum);
+        sum = fma(in_texel.yyyy, texelFetch(kernel_in, ivec2(kxs.y, ky), 0), sum);
+        sum = fma(in_texel.zzzz, texelFetch(kernel_in, ivec2(kxs.z, ky), 0), sum);
+        sum = fma(in_texel.wwww, texelFetch(kernel_in, ivec2(kxs.w, ky), 0), sum);
+      }
+    }
+  }
+
+  imageStore(image_out, pos, sum);
+}

backends/vulkan/runtime/graph/ops/glsl/conv_transpose2d.yaml

@@ -0,0 +1,18 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+conv_transpose2d:
+  parameter_names_with_default_values:
+    NDIM: 3
+    DTYPE: float
+  generate_variant_forall:
+    DTYPE:
+      - VALUE: half
+        SUFFIX: half
+      - VALUE: float
+        SUFFIX: float
+  shader_variants:
+    - NAME: conv_transpose2d

backends/vulkan/runtime/graph/ops/glsl/conv_transpose2d_prepack_weights.glsl

@@ -0,0 +1,113 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#version 450 core
+
+#define PRECISION ${PRECISION}
+
+#include "indexing_utils.h"
+
+layout(std430) buffer;
+
+layout(set = 0, binding = 0, ${IMAGE_FORMAT[DTYPE]}) uniform PRECISION restrict writeonly ${IMAGE_T[2][DTYPE]} image_out;
+layout(set = 0, binding = 1) buffer  PRECISION restrict readonly Buffer {
+  ${T[DTYPE]} data[];
+}
+buffer_in;
+
+// Corresponds to {1,4,6,36} in the example below.
+layout(set = 0, binding = 2) uniform PRECISION restrict GpuSizes {
+  ivec4 data;
+}
+gpu_sizes;
+
+// Corresponds to {3,3,7,10} in the example below.
+layout(set = 0, binding = 3) uniform PRECISION restrict OriginalSizes {
+  ivec4 data;
+}
+original_sizes;
+
+// Corresponds to {8,12} in the example below.
+layout(set = 0, binding = 4) uniform PRECISION restrict PaddedSizes {
+  ivec2 data;
+}
+padded_sizes;
+
+layout(local_size_x_id = 0, local_size_y_id = 1, local_size_z_id = 2) in;
+
+/*
+ * Computes special prepacking for a 2D transpose convolution. Each shader
+ * invocation calculates the input buffer location to read into the desired
+ * texel.
+ *
+ * For details, refer to conv2d_prepack_weights.glsl which uses a similar
+ * approach. For transpose, there are slight differences to reflect the data
+ * access pattern in the shader. First, the weight tensor is flipped along the H
+ * and W dims. Second, steps 3 and 4 are slightly different so that the splits
+ * are interleaved.
+ */
+void main() {
+  const ivec3 pos = ivec3(gl_GlobalInvocationID);
+  const ivec4 coord = POS_TO_COORD_CHANNELS_PACKED(pos, gpu_sizes.data);
+
+  if (any(greaterThanEqual(coord, gpu_sizes.data))) {
+    return;
+  }
+
+  // As in usual staging shaders, map from GPU texel position to normal CPU
+  // buffer indices: (36,6) -> (4,6,36)
+  const int base_index = COORD_TO_BUFFER_IDX(coord, gpu_sizes.data);
+  const ivec4 p0 =
+      base_index + ivec4(0, 1, 2, 3) * STRIDE_CHANNELS_PACKED(gpu_sizes.data);
+
+  // Re-map the normal CPU buffer indices to special indices, through a series
+  // of mappings: reshape is a no-op to the underlying indices, so we only map
+  // for flip, pad, and permute.
+  const int Np = padded_sizes.data.y;
+  const int Cp = padded_sizes.data.x;
+  const int N = original_sizes.data.w;
+  const int C = original_sizes.data.z;
+  const int H = original_sizes.data.y;
+  const int W = original_sizes.data.x;
+
+  // Undo step 6 premute: (4,2,3,36) -> (2,4,3,36)
+  // In the following comments, a=b=c=3.
+  // Undo step 3 permute, part 1: (8,a,b,c,4) -> (8,a,c,b,4)
+  // Undo step 3 permute, part 2: (8,a,c,b,4) -> (8,c,a,b,4)
+  // Undo step 3 permute, part 3: (8,c,a,b,4) -> (8,c,a,4,b)
+  // Undo step 3 permute, part 4: (8,c,a,4,b) -> (8,c,4,a,b)
+  const ivec4 p1 = SWAP_ADJ_DIMS(p0, 4, (Cp / 4), (H * Np * W));
+  const ivec4 p2 = SWAP_ADJ_DIMS(p1, W, (Np / 4), 4);
+  const ivec4 p3 = SWAP_ADJ_DIMS(p2, H, (Np / 4), (W * 4));
+  const ivec4 p4 = SWAP_ADJ_DIMS(p3, W, 4, 1);
+  const ivec4 p5 = SWAP_ADJ_DIMS(p4, H, 4, W);
+
+  // Undo step 0 permute: (8,12,3,3) -> (12,8,3,3)
+  const ivec4 p6 = SWAP_ADJ_DIMS(p5, Cp, Np, (W * H));
+  // Undo step 0 flip: (2,3)
+  const ivec4 w = p6 % W;
+  const ivec4 h = p6 % (H * W) / W;
+  const ivec4 p7 = p6 + W - 1 - 2 * w + W * (H - 1 - 2 * h);
+
+  // Undo step 1 pad: (12,8,3,3) -> (10,7,3,3)
+  // For values in the padded region, write zero instead of buffer data.
+  const ivec4 c = p7 % (Cp * H * W) / (H * W);
+  const ivec4 n = p7 / (Cp * H * W);
+  const ivec4 p8 = p7 - n * (Cp - C) * H * W;
+  const ivec4 mask = ivec4(greaterThanEqual(c, ivec4(C))) |
+      ivec4(greaterThanEqual(n, ivec4(N)));
+
+  ${T[DTYPE]} val_x = mix(buffer_in.data[p8.x], 0, mask.x);
+  ${T[DTYPE]} val_y = mix(buffer_in.data[p8.y], 0, mask.y);
+  ${T[DTYPE]} val_z = mix(buffer_in.data[p8.z], 0, mask.z);
+  ${T[DTYPE]} val_w = mix(buffer_in.data[p8.w], 0, mask.w);
+
+  ${VEC4_T[DTYPE]} texel = ${VEC4_T[DTYPE]}(val_x, val_y, val_z, val_w);
+
+  imageStore(image_out, pos.xy, texel);
+}

backends/vulkan/runtime/graph/ops/glsl/conv_transpose2d_prepack_weights.yaml

@@ -0,0 +1,18 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+conv_transpose2d_prepack_weights:
+  parameter_names_with_default_values:
+    NDIM: 3
+    DTYPE: float
+  generate_variant_forall:
+    DTYPE:
+      - VALUE: half
+        SUFFIX: half
+      - VALUE: float
+        SUFFIX: float
+  shader_variants:
+    - NAME: conv_transpose2d_prepack_weights