shader variants with/without unrolling

Author: jeffbolznv

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

@@ -3094,9 +3094,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         uint32_t conv2d_BS_NPQ = 128;
         uint32_t conv2d_TS_K   = 8;
         uint32_t conv2d_SHMEM_PAD = 4;
+        bool conv2d_UNROLL = true;
 
         if (device->vendor_id == VK_VENDOR_ID_INTEL) {
             conv2d_SHMEM_PAD = 0;
+            conv2d_UNROLL = false;
         }
 
         switch (s) {
@@ -3141,14 +3143,24 @@ 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, 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, conv2d_SHMEM_PAD }, 1, true, use_collectives);
+        std::array<uint32_t, 3> wg_denoms = { conv2d_BS_K, conv2d_BS_NPQ, 1 };
+        std::vector<uint32_t> spec_constants = { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives, conv2d_SHMEM_PAD };
+
+        if (conv2d_UNROLL) {
+            ggml_vk_create_pipeline(
+                device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_unroll_len, conv2d_f32_unroll_data, "main", 3,
+                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+            ggml_vk_create_pipeline(
+                device, device->pipeline_conv2d_f16_f32[s], "conv2d_f16_f32", conv2d_f16_f32_unroll_len, conv2d_f16_f32_unroll_data, "main", 3,
+                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+        } else {
+            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), wg_denoms, spec_constants, 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), wg_denoms, spec_constants, 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

@@ -202,7 +202,7 @@ void main() {
             Ash[B_ly * Ash_stride + B_lx] = val;
         }
         /* Load input to B_block: (BS_CRS x BS_NPQ) */
-        [[unroll]] for (uint32_t r_offset = 0; r_offset < BS_CRS; r_offset += BrpWg) {
+        UNROLL for (uint32_t r_offset = 0; r_offset < BS_CRS; r_offset += BrpWg) {
             uint32_t B_ly          = r_offset + Br;             /* Row index of B block */
             uint32_t B_lx          = Bc;
             uint32_t NPQ_idx       = B_idx_NPQ * BS_NPQ + B_lx; /* Global NPQ index (column index of B) */
@@ -248,7 +248,7 @@ void main() {
         }
         barrier();
         if (T_y * TS_K < K) {
-            [[unroll]] for (uint32_t CRS_lidx = 0; CRS_lidx < BS_CRS; CRS_lidx++) {
+            UNROLL for (uint32_t CRS_lidx = 0; CRS_lidx < BS_CRS; CRS_lidx++) {
                 for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
                     regA[T_ly] = Ash[(T_y * TS_K + T_ly) * Ash_stride + CRS_lidx];
                 }

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

@@ -655,8 +655,11 @@ void process_shaders() {
 
     string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
 
-    string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}});
-    string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}});
+    string_to_spv("conv2d_f32_unroll", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}});
+    string_to_spv("conv2d_f16_f32_unroll", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}});
+
+    string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
+    string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
 
     string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
     string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));