pytorch · SS-JIA · Mar 26, 2026 · Mar 25, 2026 · Mar 25, 2026
@@ -45,7 +45,7 @@ void main() {
   // Compute the value for each element in the texel along the packed dim.
   VEC4_T outtex = VEC4_T(0);
   int limit = min(
-      4, outp.sizes[packed_dim] - out_tidx.data[packed_dim]);
+      4, safe_idx(outp.sizes, packed_dim) - out_tidx.data[packed_dim]);
   for (int comp = 0; comp < limit; comp++) {
     int elem_idx = out_tidx.data[0]; // W index is the linear element index
     outtex[comp] = VEC4_T(start + elem_idx * step).x;

@@ -45,7 +45,7 @@ void main() {
   VEC4_T out_texel = VEC4_T(0);
 
   int limit = min(
-      4, outp.sizes[packed_dim] - out_tidx.data[packed_dim]);
+      4, safe_idx(outp.sizes, packed_dim) - out_tidx.data[packed_dim]);
   for (int comp = 0; comp < 4; comp++) {
     if (comp >= limit) {
       break;

@@ -40,7 +40,7 @@ void main() {
 
   TensorIndex4D tidx =
       texture_pos_to_tensor4d_idx_simple(outp, pos, out_layout);
-  const int packed_dim_size = outp.sizes[packed_dim];
+  const int packed_dim_size = safe_idx(outp.sizes, packed_dim);
   int packed_idx = tidx.data[packed_dim];
 
   if (packed_idx + 3 >= packed_dim_size) {

@@ -53,7 +53,7 @@ void main() {
   VEC4_T out_texel = VEC4_T(0);
 
   int limit = min(
-      4, outp.sizes[out_packed_dim] - out_tidx.data[out_packed_dim]);
+      4, safe_idx(outp.sizes, out_packed_dim) - out_tidx.data[out_packed_dim]);
   for (int comp = 0; comp < 4; comp++) {
     TensorIndex4D input_tidx = out_tidx;
     int gather_idx = idx_texel[comp];

@@ -55,7 +55,7 @@ void main() {
   VEC4_T out_texel = VEC4_T(0);
 
   int limit = min(
-      4, outp.sizes[out_packed_dim] - out_tidx.data[out_packed_dim]);
+      4, safe_idx(outp.sizes, out_packed_dim) - out_tidx.data[out_packed_dim]);
   for (int comp = 0; comp < limit; comp++) {
     int idx = idx_texel[comp];
 

@@ -99,6 +99,14 @@ uint safe_idx(const uvec4 v, const int idx) {
   return v.w;
 }
 
+// Safe ivec3 component access via if/else chain. Same rationale as safe_idx
+// for ivec4.
+int safe_idx(const ivec3 v, const int idx) {
+  if (idx == 0) return v.x;
+  if (idx == 1) return v.y;
+  return v.z;
+}
+
 // Safe ivec4 component write via if/else chain. Companion to safe_idx for
 // cases where we need to set a component by a spec-const-derived index.
 void safe_set(inout ivec4 v, const int idx, const int val) {

@@ -42,7 +42,7 @@ void main() {
       texel_idx_to_tensor4d_idx(outp, texel_idx, outp_layout);
 
   // Bounds check on outer dimension
-  if (tidx.data[outer_dim] >= int(outp.sizes[0][outer_dim])) {
+  if (tidx.data[outer_dim] >= int(safe_idx(outp.sizes[0], outer_dim))) {
     return;
   }
 
@@ -55,7 +55,7 @@ void main() {
   int packed = 0;
   [[unroll]] for (int i = 0; i < 4; ++i) {
     const int elem_inner = tidx.data[inner_dim] + i;
-    if (elem_inner < int(outp.sizes[0][inner_dim])) {
+    if (elem_inner < int(safe_idx(outp.sizes[0], inner_dim))) {
       // Build element coordinates
       ivec4 elem = tidx.data;
       elem[inner_dim] = elem_inner;

@@ -52,7 +52,7 @@ void main() {
 
   // Tail texels may have fewer than 4 valid elements; leave extras as 0.
   const int limit =
-      min(4, outp.sizes[packed_dim] - out_tidx.data[packed_dim]);
+      min(4, safe_idx(outp.sizes, packed_dim) - out_tidx.data[packed_dim]);
 
   VEC4_T out_texel = VEC4_T(0);
 

@@ -9,97 +9,97 @@
 #version 450 core
 
 ${define_required_extensions("texture3d", DTYPE)}
-${define_explicit_type_extensions(DTYPE)}
 
 #define PRECISION ${PRECISION}
 
-#define VEC4_T ${texel_type(DTYPE)}
-#define T ${buffer_scalar_type(DTYPE)}
+#define VEC4_T ${texel_load_type(DTYPE, "texture3d")}
+#define T ${texel_load_component_type(DTYPE, "texture3d")}
 
 ${define_active_storage_type("texture3d")}
 
+#extension GL_EXT_control_flow_attributes : require
+
 layout(std430) buffer;
 
-#include "indexing_utils.h"
+#include "common.glslh"
+#include "indexing.glslh"
 
 ${layout_declare_tensor(B, "w", "t_out", DTYPE, "texture3d")}
 ${layout_declare_tensor(B, "r", "t_in", DTYPE, "texture3d")}
 
+${layout_declare_ubo(B, "TextureMetadata", "outp")}
+${layout_declare_ubo(B, "TextureMetadata", "inp")}
+
 layout(push_constant) uniform restrict Block {
-  ivec4 out_sizes;
-  ivec4 in_sizes;
-  ivec4 permute_dims; // Permutation mapping: permute_dims[i] = j means output dim i comes from input dim j
+  ivec4 permute_dims;
 };
 
-${layout_declare_spec_const(C, "int", "out_layout", "DEFAULT_LAYOUT")}
-const lowp ivec4 out_axis_map = unhash_axis_map(out_layout);
-const lowp int out_packed_dim = unhash_packed_dim(out_layout);
-
-${layout_declare_spec_const(C, "int", "in_layout", "DEFAULT_LAYOUT")}
-const lowp ivec4 in_axis_map = unhash_axis_map(in_layout);
-const lowp int in_packed_dim = unhash_packed_dim(in_layout);
+${layout_declare_spec_const(C, "int", "out_layout", "CONTIG_LAYOUT_INT")}
+${layout_declare_spec_const(C, "int", "in_layout", "CONTIG_LAYOUT_INT")}
+const int out_packed_dim = get_packed_dim(out_layout);
+const int in_packed_dim = get_packed_dim(in_layout);
 
 layout(local_size_x_id = 0, local_size_y_id = 1, local_size_z_id = 2) in;
 
-// Convert output tensor index to input tensor index based on permutation
+// Convert output tensor index to input tensor index based on permutation.
+// permute_dims[i] = j means output dim i comes from input dim j.
+// We write: in_tidx[permute_dims.{x,y,z,w}] = out_tidx.{x,y,z,w}
+// This uses literal component access on the push constant (safe) and dynamic
+// indexing into the local in_tidx variable (also safe).
 ivec4 out_tidx_to_in_tidx(const ivec4 out_tidx) {
   ivec4 in_tidx;
-
-  // Apply the permutation mapping: in_tidx[permute_dims[i]] = out_tidx[i]
   in_tidx[permute_dims.x] = out_tidx.x;
   in_tidx[permute_dims.y] = out_tidx.y;
   in_tidx[permute_dims.z] = out_tidx.z;
   in_tidx[permute_dims.w] = out_tidx.w;
-
   return in_tidx;
 }
 
-// Check if we can use the fast path where texels from the input tensor can be
-// copied directly into the output tensor. This occurs when the packed dimension
-// is preserved in the permutation, i.e. reading a texel from the output tensor
-// produces 4 texels along the same dimension as reading a texel from the input
-// tensor.
-bool can_use_fast_path() {
-  // Fast path is possible when the packed dimension is preserved in the permutation
-  // This means permute_dims[out_packed_dim] == in_packed_dim
-  return permute_dims[out_packed_dim] == in_packed_dim;
-}
-
 void main() {
-  const ivec3 lpos = ivec3(gl_GlobalInvocationID);
-  ivec4 out_tidx = lpos_to_tidx(lpos, out_sizes, out_axis_map.w, out_packed_dim);
+  const ivec3 out_pos = ivec3(gl_GlobalInvocationID);
 
-  if (any(greaterThanEqual(out_tidx, out_sizes))) {
+  if (out_of_bounds(out_pos, outp)) {
     return;
   }
 
-  if (can_use_fast_path()) {
+  TensorIndex4D out_tidx =
+      texture_pos_to_tensor4d_idx_simple(outp, out_pos, out_layout);
+
+  // Check if packed dimension is preserved in the permutation. Use safe_idx
+  // to avoid dynamic indexing of push constant with spec-const-derived index.
+  const bool fast_path =
+      safe_idx(permute_dims, out_packed_dim) == in_packed_dim;
+
+  if (fast_path) {
     // Fast path: packed dimension is preserved, so we can copy texels directly
-    ivec4 in_tidx = out_tidx_to_in_tidx(out_tidx);
-    ivec3 in_pos = tidx_to_pos(in_tidx, in_sizes, in_axis_map, in_packed_dim);
-    VEC4_T in_texel = VEC4_T(load_texel(t_in, in_pos));
+    ivec4 in_tidx_data = out_tidx_to_in_tidx(out_tidx.data);
+    TensorIndex4D in_tidx;
+    in_tidx.data = in_tidx_data;
 
-    write_texel_lpos(t_out, lpos, in_texel, out_axis_map);
-  }
-  else {
+    ivec3 in_pos =
+        tensor4d_idx_to_texel_pos_simple(inp, in_tidx, in_layout);
+    VEC4_T in_texel = texelFetch(t_in, in_pos, 0);
+
+    imageStore(t_out, out_pos, in_texel);
+  } else {
     // Slow path: packed dimension is not preserved, so each element of the
-    // output texel may be "sourced" from a different texel in the input tensor.
-    // Therefore each output texel element is processed individually.
+    // output texel may come from a different texel in the input tensor.
     VEC4_T out_texel = VEC4_T(0);
 
-    for (int texel_i = 0; texel_i < 4; ++texel_i) {
-      ivec4 in_tidx = out_tidx_to_in_tidx(out_tidx);
-      ivec3 in_pos = tidx_to_pos(in_tidx, in_sizes, in_axis_map, in_packed_dim);
-      int element_idx = in_tidx[in_packed_dim] % 4;
+    for (int comp = 0; comp < 4; comp++) {
+      ivec4 in_tidx_data = out_tidx_to_in_tidx(out_tidx.data);
+      TensorIndex4D in_tidx;
+      in_tidx.data = in_tidx_data;
 
-      VEC4_T in_texel = VEC4_T(load_texel(t_in, in_pos));
-      T selected_value = T(in_texel[element_idx]);
+      TextureElementIndex in_elem =
+          tensor4d_idx_to_texture_element_idx_simple(inp, in_tidx, in_layout);
 
-      out_texel[texel_i] = selected_value;
+      VEC4_T in_texel = texelFetch(t_in, in_elem.pos, 0);
+      out_texel[comp] = in_texel[in_elem.comp];
 
-      out_tidx[out_packed_dim]++;
+      out_tidx.data[out_packed_dim]++;
     }
 
-    write_texel_lpos(t_out, lpos, out_texel, out_axis_map);
+    imageStore(t_out, out_pos, out_texel);
   }
 }
@@ -43,6 +43,7 @@ layout(constant_id = 5) const int group_dim = 1;
 shared vec4 shared_vecs[MAX_NTHREADS];
 
 #include "indexing_utils.h"
+#include "indexing.glslh"
 
 int tid_to_smi(const ivec2 tid) {
   return tid.x + tid.y * NWORKERS;
@@ -95,7 +96,7 @@ void reduce_nonpacked_dim(const ivec2 tid, ivec3 scan_pos) {
   scan_pos[reduce_dim] = tid.x;
   // Partially accumulate over elements i, i + NWORKERS, i + 2*NWORKERS, ... of
   // the reduction row
-  for (int i = tid.x; i < tin_sizes[reduce_dim];
+  for (int i = tid.x; i < safe_idx(tin_sizes, reduce_dim);
        i += NWORKERS, scan_pos[reduce_dim] += NWORKERS) {
     accum = UPDATE_ACCUM(accum, load_texel(tin, scan_pos));
   }
@@ -115,11 +116,11 @@ void reduce_nonpacked_dim(const ivec2 tid, ivec3 scan_pos) {
 
     // Determine if there are any padding elements in the final texel of the
     // packed dimension
-    const int nspill = mod4(tin_sizes[packed_dim]);
+    const int nspill = mod4(safe_idx(tin_sizes, packed_dim));
     // Detect if this thread is working on the final texels of the packed
     // dimension, which may have padding elements
     const bool is_last_texel =
-        scan_pos[packed_dim] == (tin_limits[packed_dim] - 1);
+        scan_pos[packed_dim] == (safe_idx(tin_limits, packed_dim) - 1);
 
     // Explicitly set padding elements to 0
     if (is_last_texel && nspill > 0) {
@@ -145,10 +146,10 @@ void reduce_packed_dim(const ivec2 tid, ivec3 scan_pos) {
   const int smi = tid_to_smi(tid);
 
   // Number of non-padding elements in the last texel in the reduction row
-  const int nspill = mod4(tin_sizes[packed_dim]);
+  const int nspill = mod4(safe_idx(tin_sizes, packed_dim));
   // Only reduce up to the last "complete" texel. The last texel will need to be
   // handled specially if it has padding elements.
-  const int reduce_len = tin_sizes[packed_dim] - nspill;
+  const int reduce_len = safe_idx(tin_sizes, packed_dim) - nspill;
 
   scan_pos[reduce_dim] = 0;
   vec4 accum = INIT_ACCUM(vec4(load_texel(tin, scan_pos).x));
@@ -163,7 +164,7 @@ void reduce_packed_dim(const ivec2 tid, ivec3 scan_pos) {
   // For the last texel in the dim, if there are padding elements then each
   // element of the texel needs to be processed individually such that the
   // padding elements are ignored
-  if (scan_pos[reduce_dim] == tin_limits[reduce_dim] - 1 && nspill > 0) {
+  if (scan_pos[reduce_dim] == safe_idx(tin_limits, reduce_dim) - 1 && nspill > 0) {
     const vec4 intex = load_texel(tin, scan_pos);
     for (int i = 0; i < nspill; i++) {
       accum.x = UPDATE_ACCUM(accum.x, intex[i]);

@@ -44,6 +44,7 @@ layout(constant_id = 6) const int group_dim = 2;
 shared vec4 shared_vecs[MAX_NTHREADS];
 
 #include "indexing_utils.h"
+#include "indexing.glslh"
 
 int tid_to_smi(const ivec2 tid) {
   return tid.x + tid.y * NWORKERS;
@@ -68,12 +69,12 @@ void reduce_2d_non_packed_dim(const ivec2 tid, ivec3 scan_pos) {
 
   // First dimension reduction
   scan_pos[reduce_dim1] = tid.x;
-  for (int i = tid.x; i < tin_sizes[reduce_dim1]; 
+  for (int i = tid.x; i < safe_idx(tin_sizes, reduce_dim1);
        i += NWORKERS, scan_pos[reduce_dim1] += NWORKERS) {
 
     // Second dimension reduction
     scan_pos[reduce_dim2] = 0;
-    for (int j = 0; j < tin_sizes[reduce_dim2]; j++, scan_pos[reduce_dim2]++) {
+    for (int j = 0; j < safe_idx(tin_sizes, reduce_dim2); j++, scan_pos[reduce_dim2]++) {
       accum = UPDATE_ACCUM(accum, load_texel(tin, scan_pos));
     }
   }
@@ -93,11 +94,11 @@ void reduce_2d_non_packed_dim(const ivec2 tid, ivec3 scan_pos) {
 
     // Determine if there are any padding elements in the final texel of the
     // packed dimension
-    const int nspill = mod4(tin_sizes[packed_dim]);
+    const int nspill = mod4(safe_idx(tin_sizes, packed_dim));
     // Detect if this thread is working on the final texels of the packed
     // dimension, which may have padding elements
-    const bool is_last_texel = 
-        scan_pos[packed_dim] == (tin_limits[packed_dim] - 1);
+    const bool is_last_texel =
+        scan_pos[packed_dim] == (safe_idx(tin_limits, packed_dim) - 1);
 
     // Explicitly set padding elements to 0
     if (is_last_texel && nspill > 0) {

@@ -47,7 +47,7 @@ void main() {
   VEC4_T out_texel = VEC4_T(0);
 
   const int limit = min(
-      4, out_meta.sizes[packed_dim] - out_tidx.data[packed_dim]);
+      4, safe_idx(out_meta.sizes, packed_dim) - out_tidx.data[packed_dim]);
   for (int comp = 0; comp < limit; comp++) {
     TensorIndex4D in_tidx = out_tidx;
     in_tidx.data = ivec4(

@@ -69,7 +69,7 @@ TensorIndex4D out_tidx_to_in_tidx(const TensorIndex4D out_tidx) {
 
   int adjusted_index = index;
   if (index < 0) {
-    adjusted_index = index + inp.sizes[selected_dim];
+    adjusted_index = index + safe_idx(inp.sizes, selected_dim);
   }
 
   // Handle different dimensions for selection

@@ -56,7 +56,7 @@ TensorIndex4D out_tidx_to_in_tidx(const TensorIndex4D out_tidx) {
 
   int adjusted_start = start;
   if (start < 0) {
-    adjusted_start = start + inp.sizes[selected_dim];
+    adjusted_start = start + safe_idx(inp.sizes, selected_dim);
   }
 
   in_tidx.data[selected_dim] = adjusted_start + out_tidx.data[selected_dim] * step;