vulkan: implement GGML_OP_SOFTMAX_BACK

Author: remyoudompheng

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

@@ -252,6 +252,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_diag_mask_inf_f32;
     vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16;
     vk_pipeline pipeline_soft_max_f32_wg512, pipeline_soft_max_f32_f16_wg512;
+    vk_pipeline pipeline_soft_max_back_f32;
     vk_pipeline pipeline_rope_norm_f32, pipeline_rope_norm_f16;
     vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16;
     vk_pipeline pipeline_rope_multi_f32, pipeline_rope_multi_f16;
@@ -2195,6 +2196,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_wg512, "soft_max_f32_wg512", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
     ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16, "soft_max_f32_f16", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
     ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16_wg512, "soft_max_f32_f16_wg512", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_back_f32, "soft_max_back_f32", soft_max_back_f32_len, soft_max_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32, "rope_norm_f32", rope_norm_f32_len, rope_norm_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
@@ -5359,6 +5361,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return src0->ne[0] > 1024 ? ctx->device->pipeline_soft_max_f32_f16_wg512 : ctx->device->pipeline_soft_max_f32_f16;
         }
         return nullptr;
+    case GGML_OP_SOFT_MAX_BACK:
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_soft_max_back_f32;
+        }
+        return nullptr;
     case GGML_OP_ROPE:
     case GGML_OP_ROPE_BACK:
         {
@@ -5690,6 +5697,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_RMS_NORM:
     case GGML_OP_RMS_NORM_BACK:
     case GGML_OP_SOFT_MAX:
+    case GGML_OP_SOFT_MAX_BACK:
     case GGML_OP_SUM_ROWS:
     case GGML_OP_ARGMAX:
         {
@@ -6397,6 +6405,11 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
     }, dryrun);
 }
 
+static void ggml_vk_soft_max_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    float * op_params = (float *)dst->op_params;
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SOFT_MAX_BACK, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], op_params[1] }, dryrun);
+}
+
 static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool backprop, bool dryrun = false) {
     const int n_dims        = ((int32_t *) dst->op_params)[1];
     const int mode          = ((int32_t *) dst->op_params)[2];
@@ -7353,6 +7366,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_RMS_NORM_BACK:
     case GGML_OP_DIAG_MASK_INF:
     case GGML_OP_SOFT_MAX:
+    case GGML_OP_SOFT_MAX_BACK:
     case GGML_OP_ROPE:
     case GGML_OP_ROPE_BACK:
     case GGML_OP_MUL_MAT:
@@ -7415,6 +7429,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         case GGML_OP_UNARY:
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
+        case GGML_OP_SOFT_MAX_BACK:
         case GGML_OP_ROPE:
         case GGML_OP_ROPE_BACK:
         case GGML_OP_ARGSORT:
@@ -7549,6 +7564,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_SOFT_MAX:
         ggml_vk_soft_max(ctx, compute_ctx, src0, src1, node, dryrun);
 
+        break;
+    case GGML_OP_SOFT_MAX_BACK:
+        ggml_vk_soft_max_back(ctx, compute_ctx, src0, src1, node, dryrun);
+
         break;
     case GGML_OP_ROPE:
         ggml_vk_rope(ctx, compute_ctx, src0, src1, src2, node, false, dryrun);
@@ -7688,6 +7707,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
     case GGML_OP_RMS_NORM_BACK:
     case GGML_OP_DIAG_MASK_INF:
     case GGML_OP_SOFT_MAX:
+    case GGML_OP_SOFT_MAX_BACK:
     case GGML_OP_ROPE:
     case GGML_OP_ROPE_BACK:
     case GGML_OP_RESHAPE:
@@ -8636,6 +8656,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_PAD:
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
+        case GGML_OP_SOFT_MAX_BACK:
         case GGML_OP_ARGSORT:
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
@@ -9038,6 +9059,8 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
         } else {
             tensor_clone = ggml_soft_max(ggml_ctx, src_clone[0]);
         }
+    } else if (tensor->op == GGML_OP_SOFT_MAX_BACK) {
+        tensor_clone = ggml_soft_max_ext_back(ggml_ctx, src_clone[0], src_clone[1], ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]);
     } else if (tensor->op == GGML_OP_DIAG_MASK_INF) {
         tensor_clone = ggml_diag_mask_inf(ggml_ctx, src_clone[0], *(int *)tensor->op_params);
     } else if (tensor->op == GGML_OP_ROPE || tensor->op == GGML_OP_ROPE_BACK) {

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

@@ -0,0 +1,50 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+
+#include "generic_head.comp"
+#include "types.comp"
+
+layout(constant_id = 0) const uint BLOCK_SIZE = 32;
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+// In this shader Y = softmax(X) and X is not provided as input.
+
+layout (binding = 0) readonly buffer G {A_TYPE data_g[];};
+layout (binding = 1) readonly buffer Y {B_TYPE data_y[];};
+layout (binding = 2) buffer D {D_TYPE data_d[];};
+
+shared FLOAT_TYPE sum_yg[BLOCK_SIZE];
+
+void main() {
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
+    const uint tid = gl_LocalInvocationID.x;
+
+    FLOAT_TYPE scale = p.param1;
+
+    // partial sums for thread in warp
+    sum_yg[tid] = FLOAT_TYPE(0.0f);
+
+    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
+        const FLOAT_TYPE gi = FLOAT_TYPE(data_g[row*p.KX + col]);
+        const FLOAT_TYPE yi = FLOAT_TYPE(data_y[row*p.KX + col]);
+        sum_yg[tid] += yi * gi;
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            sum_yg[tid] += sum_yg[tid + s];
+        }
+        barrier();
+    }
+
+    const FLOAT_TYPE dot_yg = sum_yg[0];
+
+    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
+        data_d[row*p.KX + col] = D_TYPE(scale
+            * (FLOAT_TYPE(data_g[row*p.KX + col]) - dot_yg)
+            * FLOAT_TYPE(data_y[row*p.KX + col]));
+    }
+}

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

@@ -487,6 +487,7 @@ void process_shaders() {
 
     string_to_spv("soft_max_f32", "soft_max.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("soft_max_f32_f16", "soft_max.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}));
+    string_to_spv("soft_max_back_f32", "soft_max_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
 
     string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});