vulkan: implement GGML_OP_RMS_NORM_BACK

Author: remyoudompheng

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

@@ -240,6 +240,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_norm_f32;
     vk_pipeline pipeline_group_norm_f32;
     vk_pipeline pipeline_rms_norm_f32;
+    vk_pipeline pipeline_rms_norm_back_f32;
     vk_pipeline pipeline_gelu_f32;
     vk_pipeline pipeline_gelu_quick_f32;
     vk_pipeline pipeline_silu_f32;
@@ -2118,6 +2119,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_rms_norm_back_f32, "rms_norm_back_f32", rms_norm_back_f32_len, rms_norm_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f32, "cpy_f32_f32", cpy_f32_f32_len, cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f16, "cpy_f32_f16", cpy_f32_f16_len, cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
@@ -5270,6 +5272,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_rms_norm_f32;
         }
         return nullptr;
+    case GGML_OP_RMS_NORM_BACK:
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_rms_norm_back_f32;
+        }
+        return nullptr;
     case GGML_OP_UNARY:
         switch (ggml_get_unary_op(dst)) {
             case GGML_UNARY_OP_SILU:
@@ -5627,6 +5634,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     switch (op) {
     case GGML_OP_NORM:
     case GGML_OP_RMS_NORM:
+    case GGML_OP_RMS_NORM_BACK:
     case GGML_OP_SOFT_MAX:
     case GGML_OP_SUM_ROWS:
         {
@@ -6144,6 +6152,11 @@ static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx,
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_RMS_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
 }
 
+static void ggml_vk_rms_norm_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_RMS_NORM_BACK, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
+}
+
 static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
 }
@@ -7117,6 +7130,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_NORM:
     case GGML_OP_GROUP_NORM:
     case GGML_OP_RMS_NORM:
+    case GGML_OP_RMS_NORM_BACK:
     case GGML_OP_DIAG_MASK_INF:
     case GGML_OP_SOFT_MAX:
     case GGML_OP_ROPE:
@@ -7170,6 +7184,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         case GGML_OP_NORM:
         case GGML_OP_GROUP_NORM:
         case GGML_OP_RMS_NORM:
+        case GGML_OP_RMS_NORM_BACK:
         case GGML_OP_UNARY:
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
@@ -7267,6 +7282,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_RMS_NORM:
         ggml_vk_rms_norm(ctx, compute_ctx, src0, node, dryrun);
 
+        break;
+    case GGML_OP_RMS_NORM_BACK:
+        ggml_vk_rms_norm_back(ctx, compute_ctx, src0, src1, node, dryrun);
+
         break;
     case GGML_OP_UNARY:
         switch (ggml_get_unary_op(node)) {
@@ -7405,6 +7424,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
     case GGML_OP_NORM:
     case GGML_OP_GROUP_NORM:
     case GGML_OP_RMS_NORM:
+    case GGML_OP_RMS_NORM_BACK:
     case GGML_OP_DIAG_MASK_INF:
     case GGML_OP_SOFT_MAX:
     case GGML_OP_ROPE:
@@ -8327,6 +8347,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_CONCAT:
+        case GGML_OP_RMS_NORM_BACK:
         case GGML_OP_UPSCALE:
         case GGML_OP_SCALE:
         case GGML_OP_SQR:
@@ -8850,6 +8871,9 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
         tensor_clone = ggml_group_norm(ggml_ctx, src0_clone, *(int *)tensor->op_params, ((float *)tensor->op_params)[1]);
     } else if (tensor->op == GGML_OP_RMS_NORM) {
         tensor_clone = ggml_rms_norm(ggml_ctx, src0_clone, *(float *)tensor->op_params);
+    } else if (tensor->op == GGML_OP_RMS_NORM_BACK) {
+        const float eps = ((float *) tensor->op_params)[0];
+        tensor_clone = ggml_rms_norm_back(ggml_ctx, src0_clone, src1_clone, eps);
     } else if (tensor->op == GGML_OP_SOFT_MAX) {
         if (src1 != nullptr) {
             tensor_clone = ggml_soft_max_ext(ggml_ctx, src0_clone, src1_clone, ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]);

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

@@ -0,0 +1,55 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+#define BLOCK_SIZE 512
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer G {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer X {B_TYPE data_b[];};
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+
+shared FLOAT_TYPE sum_xx[BLOCK_SIZE];
+shared FLOAT_TYPE sum_xg[BLOCK_SIZE];
+
+void main() {
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
+    const uint tid = gl_LocalInvocationID.x;
+
+    // Compute derivative of x[i]/norm(x) = g[i]/norm(x) - x[i] dot(x,g)/KX / norm(x)^1.5
+
+    // partial sums for thread in warp
+    sum_xx[tid] = FLOAT_TYPE(0.0f);
+    sum_xg[tid] = FLOAT_TYPE(0.0f);
+
+    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
+        const FLOAT_TYPE gi = FLOAT_TYPE(data_a[row*p.KX + col]);
+        const FLOAT_TYPE xi = FLOAT_TYPE(data_b[row*p.KX + col]);
+        sum_xx[tid] += xi * xi;
+        sum_xg[tid] += xi * gi;
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            sum_xx[tid] += sum_xx[tid + s];
+            sum_xg[tid] += sum_xg[tid + s];
+        }
+        barrier();
+    }
+
+    const FLOAT_TYPE eps = FLOAT_TYPE(p.param1);
+    const FLOAT_TYPE mean = sum_xx[0] / FLOAT_TYPE(p.KX);
+    const FLOAT_TYPE scale_g = inversesqrt(mean + eps);
+    const FLOAT_TYPE scale_x = -scale_g * sum_xg[0] / (sum_xx[0] + FLOAT_TYPE(p.KX) * eps);
+
+    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
+        data_d[row*p.KX + col] = D_TYPE(
+            scale_g * FLOAT_TYPE(data_a[row*p.KX + col]) +
+            scale_x * FLOAT_TYPE(data_b[row*p.KX + col]));
+    }
+}

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

@@ -427,6 +427,7 @@ void process_shaders() {
     string_to_spv("norm_f32", "norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("group_norm_f32", "group_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    string_to_spv("rms_norm_back_f32", "rms_norm_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
 
     string_to_spv("cpy_f32_f32", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("cpy_f32_f16", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});