vulkan: support all backend tests for SQR/SQRT/SIN/COS/CLAMP/LEAKY_RELU/NORM (#24582)

* vulkan: make SQR/SQRT/SIN/COS/CLAMP/LEAKY_RELU use unary.comp

* vulkan: make NORM support noncontig

* add noncontiguous row test cases for norm/l2_norm, handle this in the CPU backend and l2_norm.comp

* fix supports_op for cuda and webgpu

Author: jeffbolznv

ggml/src/ggml-cpu/ops.cpp

@@ -3688,8 +3688,6 @@ static void ggml_compute_forward_norm_f32(
 
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    GGML_ASSERT(src0->nb[0] == sizeof(float));
-
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -3703,25 +3701,49 @@ static void ggml_compute_forward_norm_f32(
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
             for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
-                const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+                const char * x = (const char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
+                char * y = (char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3;
 
-                float sum = 0.0;
-                ggml_vec_sum_f32(ne00, &sum, x);
-                float mean = sum/ne00;
+                if (nb00 == sizeof(float) && nb0 == sizeof(float)) {
+                    const float * xf = (const float *) x;
 
-                float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
-                float variance = 0;
+                    float sum = 0.0;
+                    ggml_vec_sum_f32(ne00, &sum, xf);
+                    float mean = sum/ne00;
+
+                    float * yf = (float *) y;
+                    float variance = 0;
 
 #ifdef GGML_USE_ACCELERATE
-                mean = -mean;
-                vDSP_vsadd(x, 1, &mean, y, 1, ne00);
-                vDSP_measqv(y, 1, &variance, ne00);
+                    mean = -mean;
+                    vDSP_vsadd(xf, 1, &mean, yf, 1, ne00);
+                    vDSP_measqv(yf, 1, &variance, ne00);
 #else
-                variance = ggml_vec_cvar_f32(ne00, y, x, mean);
+                    variance = ggml_vec_cvar_f32(ne00, yf, xf, mean);
 #endif //GGML_USE_ACCELERATE
 
-                const float scale = 1.0f/sqrtf(variance + eps);
-                ggml_vec_scale_f32(ne00, y, scale);
+                    const float scale = 1.0f/sqrtf(variance + eps);
+                    ggml_vec_scale_f32(ne00, yf, scale);
+                } else {
+                    float sum = 0.0;
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        sum += *(const float *) (x + i00*nb00);
+                    }
+                    const float mean = sum/ne00;
+
+                    float variance = 0.0f;
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        const float v = *(const float *) (x + i00*nb00) - mean;
+                        *(float *) (y + i00*nb0) = v;
+                        variance += v * v;
+                    }
+                    variance /= ne00;
+
+                    const float scale = 1.0f/sqrtf(variance + eps);
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        *(float *) (y + i00*nb0) *= scale;
+                    }
+                }
             }
         }
     }
@@ -4142,8 +4164,6 @@ static void ggml_compute_forward_l2_norm_f32(
 
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    GGML_ASSERT(src0->nb[0] == sizeof(float));
-
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -4158,20 +4178,27 @@ static void ggml_compute_forward_l2_norm_f32(
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
             for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
-                const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+                const char * x = (const char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
 
                 ggml_float sum = 0.0;
                 for (int64_t i00 = 0; i00 < ne00; i00++) {
-                    sum += (ggml_float)(x[i00] * x[i00]);
+                    const float xi = *(const float *) (x + i00*nb00);
+                    sum += (ggml_float)(xi * xi);
                 }
 
-                float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
-
-                memcpy(y, x, ne00 * sizeof(float));
-
                 const float scale = 1.0f/fmaxf(sqrtf(sum), eps);
 
-                ggml_vec_scale_f32(ne00, y, scale);
+                char * y = (char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3;
+
+                if (nb00 == sizeof(float) && nb0 == sizeof(float)) {
+                    memcpy(y, x, ne00 * sizeof(float));
+                    ggml_vec_scale_f32(ne00, (float *) y, scale);
+                } else {
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        const float xi = *(const float *) (x + i00*nb00);
+                        *(float *) (y + i00*nb0) = xi * scale;
+                    }
+                }
             }
         }
     }

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -5334,7 +5334,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_NORM:
         case GGML_OP_RMS_NORM:
         case GGML_OP_L2_NORM:
-            return true;
+            return ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_RMS_NORM_BACK:
             return ggml_is_contiguous(op->src[0]);
             break;

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

@@ -816,14 +816,10 @@ struct vk_device_struct {
     vk_pipeline pipeline_concat_i8, pipeline_concat_i16, pipeline_concat_i32, pipeline_concat_i64;
     vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32, pipeline_upscale_bicubic_f32, pipeline_upscale_bilinear_antialias_f32;
     vk_pipeline pipeline_scale_f32;
-    vk_pipeline pipeline_sqr_f32;
-    vk_pipeline pipeline_sqrt_f32;
-    vk_pipeline pipeline_sin_f32;
-    vk_pipeline pipeline_cos_f32;
     vk_pipeline pipeline_log[2];
     vk_pipeline pipeline_tri[2];
     vk_pipeline pipeline_diag[2];
-    vk_pipeline pipeline_clamp_f32;
+    vk_pipeline pipeline_clamp[2];
     vk_pipeline pipeline_pad_f32;
     vk_pipeline pipeline_roll_f32;
     vk_pipeline pipeline_repeat_i32, pipeline_repeat_back_f32;
@@ -855,6 +851,10 @@ struct vk_device_struct {
     vk_pipeline pipeline_gelu_quick[2];
     vk_pipeline pipeline_silu[2];
     vk_pipeline pipeline_relu[2];
+    vk_pipeline pipeline_sqr[2];
+    vk_pipeline pipeline_sqrt[2];
+    vk_pipeline pipeline_sin[2];
+    vk_pipeline pipeline_cos[2];
     vk_pipeline pipeline_xielu[2];
     vk_pipeline pipeline_neg[2];
     vk_pipeline pipeline_tanh[2];
@@ -886,7 +886,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_geglu_erf[2];
     vk_pipeline pipeline_geglu_quick[2];
 
-    vk_pipeline pipeline_leaky_relu_f32;
+    vk_pipeline pipeline_leaky_relu[2];
     vk_pipeline pipeline_silu_back_f32;
     vk_pipeline pipeline_diag_mask_inf_f32;
     vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16;
@@ -4972,7 +4972,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     }
     ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_nc_push_constants), {1, 1, 1}, {}, 1);
 
-    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_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_unary_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", 4, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {0, 0}, 1, true);
@@ -5092,11 +5092,6 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
 
     ggml_vk_create_pipeline(device, device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_sqr_f32, "sqr_f32", sqr_f32_len, sqr_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_sqrt_f32, "sqrt_f32", sqrt_f32_len, sqrt_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_sin_f32, "sin_f32", sin_f32_len, sin_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_cos_f32, "cos_f32", cos_f32_len, cos_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-
     ggml_vk_create_pipeline(device, device->pipeline_log[0], "log_f32", log_f32_len, log_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_log[1], "log_f16", log_f16_len, log_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
@@ -5106,8 +5101,6 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     ggml_vk_create_pipeline(device, device->pipeline_diag[0], "diag_f32", diag_f32_len, diag_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_diag[1], "diag_f16", diag_f16_len, diag_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_clamp_f32, "clamp_f32", clamp_f32_len, clamp_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-
     ggml_vk_create_pipeline(device, device->pipeline_pad_f32, "pad_f32", pad_f32_len, pad_f32_data, "main", 2, sizeof(vk_op_pad_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_roll_f32, "roll_f32", roll_f32_len, roll_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
@@ -5127,6 +5120,12 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     CREATE_UNARY(gelu_quick)
     CREATE_UNARY(silu)
     CREATE_UNARY(relu)
+    CREATE_UNARY(sqr)
+    CREATE_UNARY(sqrt)
+    CREATE_UNARY(sin)
+    CREATE_UNARY(cos)
+    CREATE_UNARY(clamp)
+    CREATE_UNARY(leaky_relu)
     CREATE_UNARY(xielu)
     CREATE_UNARY(neg)
     CREATE_UNARY(tanh)
@@ -5166,7 +5165,6 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     CREATE_GLU(geglu_quick)
 #undef CREATE_GLU
 
-    ggml_vk_create_pipeline(device, device->pipeline_leaky_relu_f32, "leaky_relu_f32", leaky_relu_f32_len, leaky_relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_silu_back_f32, "silu_back_f32", silu_back_f32_len, silu_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {1, 512, 1}, {}, 1, true);
@@ -10521,23 +10519,27 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         }
         return nullptr;
     case GGML_OP_SQR:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_sqr_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_sqr[dst->type == GGML_TYPE_F16];
         }
         return nullptr;
     case GGML_OP_SQRT:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_sqrt_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_sqrt[dst->type == GGML_TYPE_F16];
         }
         return nullptr;
     case GGML_OP_SIN:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_sin_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_sin[dst->type == GGML_TYPE_F16];
         }
         return nullptr;
     case GGML_OP_COS:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_cos_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_cos[dst->type == GGML_TYPE_F16];
         }
         return nullptr;
     case GGML_OP_LOG:
@@ -10559,8 +10561,9 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         }
         return nullptr;
     case GGML_OP_CLAMP:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_clamp_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_clamp[dst->type == GGML_TYPE_F16];
         }
         return nullptr;
     case GGML_OP_PAD:
@@ -10928,8 +10931,9 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         }
         return nullptr;
     case GGML_OP_LEAKY_RELU:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_leaky_relu_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_leaky_relu[dst->type == GGML_TYPE_F16];
         }
         return nullptr;
     case GGML_OP_CONV_2D:
@@ -11431,6 +11435,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_TRI:
     case GGML_OP_DIAG:
     case GGML_OP_CLAMP:
+    case GGML_OP_LEAKY_RELU:
     case GGML_OP_PAD:
     case GGML_OP_ROLL:
     case GGML_OP_REPEAT:
@@ -12297,8 +12302,10 @@ static void ggml_vk_silu_back(ggml_backend_vk_context * ctx, vk_context& subctx,
 
 static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     float * op_params = (float *)dst->op_params;
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst);
+    p.param1 = op_params[0];
 
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f, 0.0f, 0.0f });
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_NORM, std::move(p));
 }
 
 static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
@@ -13399,7 +13406,10 @@ static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx
 
 static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     const float * op_params = (const float *)dst->op_params;
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f, 0.0f, 0.0f });
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst);
+    p.param1 = op_params[0];
+
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, std::move(p));
 }
 
 #ifdef GGML_VULKAN_RUN_TESTS
@@ -17325,12 +17335,11 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_TRANSPOSE:
         case GGML_OP_RMS_NORM:
             return true;
-        case GGML_OP_NORM:
         case GGML_OP_GROUP_NORM:
             return ggml_is_contiguous(op->src[0]);
+        case GGML_OP_NORM:
         case GGML_OP_L2_NORM:
-            return ggml_is_contiguous_rows(op->src[0]) &&
-                   op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
         case GGML_OP_ADD:
         case GGML_OP_SUB:
         case GGML_OP_MUL:
@@ -17349,8 +17358,9 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_SIN:
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
-            return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_LEAKY_RELU:
+            return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
+                   op->type == op->src[0]->type;
         case GGML_OP_OPT_STEP_ADAMW:
         case GGML_OP_OPT_STEP_SGD:
             return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;

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

@@ -14,16 +14,13 @@ void main() {
     const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint tid = gl_LocalInvocationID.x;
 
-    const uint i3 = row / (p.ne11 * p.ne12);
-    const uint i3_offset = i3 * p.ne12 * p.ne11;
-    const uint i2 = (row - i3_offset) / p.ne11;
-    const uint i2_offset = i2 * p.ne11;
-    const uint i1 = row - i3_offset - i2_offset;
+    const uint a_base = get_aoffset() + src0_idx(row * p.ne00);
+    const uint d_base = get_doffset() + dst_idx(row * p.ne10);
 
     sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp
 
     [[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) {
-        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0]);
+        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[a_base + i0*p.nb00]);
         sum[tid] += xi * xi;
     }
 
@@ -39,6 +36,6 @@ void main() {
     const FLOAT_TYPE scale = 1.0f / max(sqrt(sum[0]), FLOAT_TYPE(p.param1));
 
     [[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) {
-        data_d[i3*p.nb13 + i2*p.nb12 + i1*p.nb11 + i0] = D_TYPE(scale * FLOAT_TYPE(data_a[i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0]));
+        data_d[d_base + i0*p.nb10] = D_TYPE(scale * FLOAT_TYPE(data_a[a_base + i0*p.nb00]));
     }
 }