make SHMEM_PAD a spec constant

Author: jeffbolznv

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

@@ -3140,11 +3140,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         ggml_vk_create_pipeline(
             device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_len, conv2d_f32_data, "main", 3,
             sizeof(vk_op_conv2d_push_constants), { conv2d_BS_K, conv2d_BS_NPQ, 1 },
-            { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives }, 1, true, use_collectives);
+            { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives, conv2d_SHMEM_PAD }, 1, true, use_collectives);
         ggml_vk_create_pipeline(
             device, device->pipeline_conv2d_f16_f32[s], "conv2d_f16_f32", conv2d_f16_f32_len, conv2d_f16_f32_data, "main", 3,
             sizeof(vk_op_conv2d_push_constants), { conv2d_BS_K, conv2d_BS_NPQ, 1 },
-            { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives }, 1, true, use_collectives);
+            { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives, conv2d_SHMEM_PAD }, 1, true, use_collectives);
     }
 
     ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_whcn_f32, "conv2d_dw_whcn_f32", conv2d_dw_whcn_f32_len, conv2d_dw_whcn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);

ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp

@@ -8,9 +8,6 @@
 
 #include "types.comp"
 
-// Make spec constant
-#define SHMEM_PAD 4
-
 // shape notation: [dim(N), ..., dim(0)] -- stride(dim(j)) >= stride(dim(i)) if i > j
 layout(binding = 0) readonly buffer A {
     A_TYPE knl_data[];
@@ -76,6 +73,7 @@ layout(constant_id = 3) const uint BS_NPQ          = 128;
 // Thread-tile sizes
 layout(constant_id = 4) const uint TS_K            = 8;
 layout(constant_id = 5) const uint use_collectives = 1;
+layout(constant_id = 6) const uint SHMEM_PAD       = 4;
 
 uint32_t       tid     = gl_LocalInvocationID.x;
 const uint32_t WG_SIZE = gl_WorkGroupSize.x;