vulkan: Optimize soft_max (llama/10301)

* vulkan: Optimize soft_max

Large soft_max could already saturate memory, but small/medium sizes were
pretty slow. The bulk of the gains for them comes from using a smaller
workgroup size, and making the workgroup size match the subgroup size also
makes the barriers much cheaper.

Cache some values in locals to avoid refetching/recomputing. And stamp
out a few "template instantiations" so smaller cases will fully unroll.

Add a missing early return for OOB rows. This happens when there are more
than 512 rows and the dispatch is 512 x H.

* vulkan: Further soft_max optimizations

Restore the workgroup size of 512 case, use it for >1024.

Use unrollable loops for more iteration counts.

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

@@ -218,6 +218,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_tanh_f32;
     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_rope_norm_f32, pipeline_rope_norm_f16;
     vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16;
     vk_pipeline pipeline_argsort_f32;
@@ -388,6 +389,7 @@ struct vk_op_soft_max_push_constants {
     float m0;
     float m1;
     uint32_t n_head_log2;
+    uint32_t nrows_x;
 };
 
 struct vk_op_argsort_push_constants {
@@ -1497,8 +1499,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     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), {512, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, {}, 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}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_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_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_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_norm_f16, "rope_norm_f16", rope_norm_f16_len, rope_norm_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
@@ -3932,10 +3936,10 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);
 
         if (src0->type == GGML_TYPE_F32 && (src1 == nullptr || src1->type == GGML_TYPE_F32) && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_soft_max_f32;
+            return src0->ne[0] > 1024 ? ctx->device->pipeline_soft_max_f32_wg512 : ctx->device->pipeline_soft_max_f32;
         }
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_soft_max_f32_f16;
+            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_ROPE:
@@ -4581,6 +4585,7 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
         scale, max_bias,
         m0, m1,
         n_head_log2,
+        nrows_x,
     }, dryrun);
 }
 

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

@@ -1,6 +1,7 @@
 #version 450
 
-#extension GL_EXT_shader_16bit_storage : require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#extension GL_EXT_control_flow_attributes : enable
 
 layout (push_constant) uniform parameter
 {
@@ -11,14 +12,13 @@ layout (push_constant) uniform parameter
     float m0;
     float m1;
     uint n_head_log2;
+    uint nrows_x;
 } p;
 
 #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(constant_id = 0) const uint BLOCK_SIZE = 32;
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer Y {B_TYPE data_b[];};
@@ -26,11 +26,18 @@ layout (binding = 2) buffer D {D_TYPE data_d[];};
 
 shared FLOAT_TYPE vals[BLOCK_SIZE];
 
-void main() {
+// num_iters is the number of BLOCK_SIZE loop iterations we need to iterate
+// over all the columns. The main function tries to pass a constant here,
+// as if it were a template function, to allow unrolling.
+void soft_max(uint num_iters) {
     const uint tid = gl_LocalInvocationID.x;
     const uint rowx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint rowy = rowx % p.KY;
 
+    if (rowx >= p.nrows_x) {
+        return;
+    }
+
     float slope = 1.0f;
 
     // ALiBi
@@ -46,19 +53,37 @@ void main() {
     // Find max
     FLOAT_TYPE max_val = uintBitsToFloat(0xFF800000);
 
-    [[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
+    // Cache values while we compute the max, so we don't need to read them
+    // again when we're ready to compute exp(x-max).
+    const uint DATA_CACHE_SIZE = 16;
+    FLOAT_TYPE data_cache[DATA_CACHE_SIZE];
+
+    [[unroll]] for (uint col0 = 0, idx = 0; idx < num_iters; col0 += BLOCK_SIZE, ++idx) {
         const uint col = col0 + tid;
 
-        if (col >= p.KX) {
-            break;
+        FLOAT_TYPE a = FLOAT_TYPE(0);
+        if (col < p.KX) {
+            a = data_a[rowx * p.KX + col];
         }
 
-        max_val = max(max_val, FLOAT_TYPE(data_a[rowx * p.KX + col]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)));
+        FLOAT_TYPE b = FLOAT_TYPE(0);
+        if (p.KY > 0 && col < p.KX) {
+            b = data_b[rowy * p.KX + col];
+        }
+
+        FLOAT_TYPE v = a * p.scale + slope * b;
+
+        max_val = max(max_val, v);
+
+        if (idx < DATA_CACHE_SIZE) {
+            data_cache[idx] = v;
+        }
     }
-    vals[tid] = max_val;
 
+    // reduce across the workgroup
+    vals[tid] = max_val;
     barrier();
-    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+    [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
         if (tid < s) {
             vals[tid] = max(vals[tid], vals[tid + s]);
         }
@@ -68,39 +93,80 @@ void main() {
     max_val = vals[0];
     barrier();
 
-    // Sum up values
-    vals[tid] = FLOAT_TYPE(0.0f);
+    FLOAT_TYPE sum = FLOAT_TYPE(0.0f);
 
-    [[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
+    // Compute sum{exp(x - max)}
+    [[unroll]] for (uint col0 = 0, idx = 0; idx < num_iters; col0 += BLOCK_SIZE, ++idx) {
         const uint col = col0 + tid;
 
         if (col >= p.KX) {
             break;
         }
 
+        // compute exp(a*scale+b*slope), add it to sum, and cache the new value
+        // in data_cache if possible.
         const uint i = rowx * p.KX + col;
-        const FLOAT_TYPE val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)) - max_val);
-        vals[tid] += val;
-        data_d[i] = D_TYPE(val);
+        FLOAT_TYPE val;
+        if (idx < DATA_CACHE_SIZE) {
+            val = exp(data_cache[idx] - max_val);
+        } else {
+            val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)) - max_val);
+        }
+        sum += val;
+        if (idx < DATA_CACHE_SIZE) {
+            data_cache[idx] = val;
+        } else {
+            data_d[i] = D_TYPE(val);
+        }
     }
 
+    // reduce across the workgroup
+    vals[tid] = sum;
     barrier();
-    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+    [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
         if (tid < s) {
             vals[tid] += vals[tid + s];
         }
         barrier();
     }
+    sum = vals[0];
 
-    const D_TYPE divisor = D_TYPE(vals[0]);
+    FLOAT_TYPE rcpdivisor = 1.0/sum;
 
-    [[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
+    [[unroll]] for (uint col0 = 0, idx = 0; idx < num_iters; col0 += BLOCK_SIZE, ++idx) {
         const uint col = col0 + tid;
 
         if (col >= p.KX) {
-            break;
+            continue;
+        }
+
+        if (idx < DATA_CACHE_SIZE) {
+            data_d[rowx*p.KX + col] = D_TYPE(data_cache[idx] * rcpdivisor);
+        } else {
+            data_d[rowx*p.KX + col] *= D_TYPE(rcpdivisor);
         }
+    }
+}
 
-        data_d[rowx*p.KX + col] /= divisor;
+void main() {
+    // instantiate the soft_max function for several different
+    // dimensions, to allow loop unrolling
+    uint num_blocks = (p.KX + BLOCK_SIZE - 1) / BLOCK_SIZE;
+    if (num_blocks > 32) {
+        soft_max(num_blocks);
+    } else if (num_blocks > 16) {
+        soft_max(32);
+    } else if (num_blocks > 8) {
+        soft_max(16);
+    } else if (num_blocks > 4) {
+        soft_max(8);
+    } else if (num_blocks == 4) {
+        soft_max(4);
+    } else if (num_blocks == 3) {
+        soft_max(3);
+    } else if (num_blocks == 2) {
+        soft_max(2);
+    } else if (num_blocks == 1) {
+        soft_max(1);
     }
 }