Option to JIT steel gemm / conv (#1139)

Author: awni

mlx/backend/metal/CMakeLists.txt

@@ -1,4 +1,4 @@
-function(make_jit_source SRC_NAME)
+function(make_jit_source SRC_FILE)
   # This function takes a metal header file,
   # runs the C preprocessesor on it, and makes
   # the processed contents available as a string in a C++ function
@@ -9,17 +9,18 @@ function(make_jit_source SRC_NAME)
   #
   # Additional arguments to this function are treated as dependencies
   # in the Cmake build system.
+  get_filename_component(SRC_NAME ${SRC_FILE} NAME)
   add_custom_command(
     OUTPUT  jit/${SRC_NAME}.cpp
     COMMAND /bin/bash
               ${CMAKE_CURRENT_SOURCE_DIR}/make_compiled_preamble.sh
               ${CMAKE_CURRENT_BINARY_DIR}/jit
               ${CMAKE_C_COMPILER}
               ${PROJECT_SOURCE_DIR}
-              ${SRC_NAME}
+              ${SRC_FILE}
               "-D${MLX_METAL_VERSION}"
     DEPENDS make_compiled_preamble.sh
-            kernels/${SRC_NAME}.h
+            kernels/${SRC_FILE}.h
             ${ARGN}
   )
   add_custom_target(${SRC_NAME} DEPENDS jit/${SRC_NAME}.cpp)
@@ -73,6 +74,39 @@ if (MLX_METAL_JIT)
     kernels/reduction/reduce_col.h
     kernels/reduction/reduce_row.h
   )
+  make_jit_source(
+    steel/gemm/gemm
+    kernels/steel/utils.h
+    kernels/steel/gemm/loader.h
+    kernels/steel/gemm/mma.h
+    kernels/steel/gemm/params.h
+    kernels/steel/gemm/transforms.h
+  )
+  make_jit_source(steel/gemm/kernels/steel_gemm_fused)
+  make_jit_source(
+    steel/gemm/kernels/steel_gemm_masked
+    kernels/steel/defines.h
+  )
+  make_jit_source(steel/gemm/kernels/steel_gemm_splitk)
+  make_jit_source(
+    steel/conv/conv
+    kernels/steel/utils.h
+    kernels/steel/defines.h
+    kernels/steel/gemm/mma.h
+    kernels/steel/gemm/transforms.h
+    kernels/steel/conv/params.h
+    kernels/steel/conv/loader.h
+    kernels/steel/conv/loaders/loader_channel_l.h
+    kernels/steel/conv/loaders/loader_channel_n.h
+  )
+  make_jit_source(
+    steel/conv/kernels/steel_conv
+  )
+  make_jit_source(
+    steel/conv/kernels/steel_conv_general
+    kernels/steel/defines.h
+    kernels/steel/conv/loaders/loader_general.h
+  )
 else()
   target_sources(
     mlx

mlx/backend/metal/conv.cpp

@@ -7,6 +7,7 @@
 
 #include "mlx/backend/metal/copy.h"
 #include "mlx/backend/metal/device.h"
+#include "mlx/backend/metal/kernels.h"
 #include "mlx/backend/metal/kernels/defines.h"
 #include "mlx/backend/metal/kernels/steel/conv/params.h"
 #include "mlx/backend/metal/matmul.h"
@@ -335,7 +336,17 @@ void implicit_gemm_conv_2D_gpu(
 
   // Encode and dispatch kernel
   auto& compute_encoder = d.get_command_encoder(s.index);
-  auto kernel = d.get_kernel(kname.str());
+  auto kernel = get_steel_conv_kernel(
+      d,
+      kname.str(),
+      out,
+      bm,
+      bn,
+      bk,
+      wm,
+      wn,
+      n_channel_specialization,
+      small_filter);
   compute_encoder->setComputePipelineState(kernel);
 
   // Deduce grid launch dimensions
@@ -488,7 +499,8 @@ void implicit_gemm_conv_2D_general_gpu(
 
   // Encode and dispatch kernel
   auto& compute_encoder = d.get_command_encoder(s.index);
-  auto kernel = d.get_kernel(kname.str());
+  auto kernel =
+      get_steel_conv_general_kernel(d, kname.str(), out, bm, bn, bk, wm, wn);
   compute_encoder->setComputePipelineState(kernel);
 
   // Deduce grid launch dimensions

mlx/backend/metal/jit/includes.h

@@ -23,4 +23,12 @@ const char* softmax();
 const char* sort();
 const char* reduce();
 
+const char* gemm();
+const char* steel_gemm_fused();
+const char* steel_gemm_masked();
+const char* steel_gemm_splitk();
+const char* conv();
+const char* steel_conv();
+const char* steel_conv_general();
+
 } // namespace mlx::core::metal

mlx/backend/metal/jit/steel_conv.h

@@ -0,0 +1,32 @@
+// Copyright © 2024 Apple Inc.
+
+constexpr std::string_view steel_conv_kernels = R"(
+template [[host_name("{name}")]] [[kernel]] void
+implicit_gemm_conv_2d<{itype}, {bm}, {bn}, {bk}, {wm}, {wn}, {n_channels}, {small_filter}>(
+    const device {itype}* A [[buffer(0)]],
+    const device {itype}* B [[buffer(1)]],
+    device {itype}* C [[buffer(2)]],
+    const constant MLXConvParams<2>* params [[buffer(3)]],
+    const constant ImplicitGemmConv2DParams* gemm_params [[buffer(4)]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint3 lid [[thread_position_in_threadgroup]],
+    uint simd_gid [[simdgroup_index_in_threadgroup]],
+    uint simd_lid [[thread_index_in_simdgroup]]);
+)";
+
+constexpr std::string_view steel_conv_general_kernels = R"(
+template [[host_name("{name}")]] [[kernel]] void
+    implicit_gemm_conv_2d_general<{itype}, {bm}, {bn}, {bk}, {wm}, {wn}>(
+        const device {itype}* A [[buffer(0)]],
+        const device {itype}* B [[buffer(1)]],
+        device {itype}* C [[buffer(2)]],
+        const constant MLXConvParams<2>* params [[buffer(3)]],
+        const constant ImplicitGemmConv2DParams* gemm_params [[buffer(4)]],
+        const constant Conv2DGeneralJumpParams* jump_params [[buffer(5)]],
+        const constant Conv2DGeneralBaseInfo* base_h [[buffer(6)]],
+        const constant Conv2DGeneralBaseInfo* base_w [[buffer(7)]],
+        uint3 tid [[threadgroup_position_in_grid]],
+        uint3 lid [[thread_position_in_threadgroup]],
+        uint simd_gid [[simdgroup_index_in_threadgroup]],
+        uint simd_lid [[thread_index_in_simdgroup]]);
+)";

mlx/backend/metal/jit/steel_gemm.h

@@ -0,0 +1,106 @@
+// Copyright © 2024 Apple Inc.
+
+constexpr std::string_view steel_gemm_fused_kernels = R"(
+template [[host_name("{name}")]]
+[[kernel]] void gemm<{itype}, {bm}, {bn}, {bk}, {wm}, {wn}, {trans_a}, {trans_b}, float>(
+    const device {itype} *A [[buffer(0)]],
+    const device {itype} *B [[buffer(1)]],
+    const device {itype} *C [[buffer(2), function_constant(use_out_source)]],
+    device {itype} *D [[buffer(3)]],
+    const constant GEMMParams* params [[buffer(4)]],
+    const constant GEMMAddMMParams* addmm_params [[buffer(5), function_constant(use_out_source)]],
+    const constant int* batch_shape [[buffer(6)]],
+    const constant size_t* batch_strides [[buffer(7)]],
+    const constant uint32_t* lhs_indices [[buffer(10), function_constant(do_gather)]],
+    const constant uint32_t* rhs_indices [[buffer(11), function_constant(do_gather)]],
+    const constant uint32_t* C_indices [[buffer(12), function_constant(gather_bias)]],
+    const constant int* operand_shape [[buffer(13), function_constant(do_gather)]],
+    const constant size_t* operand_strides [[buffer(14), function_constant(do_gather)]],
+    const constant packed_int3& operand_batch_ndim [[buffer(15), function_constant(do_gather)]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint3 lid [[thread_position_in_threadgroup]]);
+)";
+
+constexpr std::string_view steel_gemm_masked_kernels = R"(
+template [[host_name("{name}")]] [[kernel]] void
+block_masked_gemm<
+    {itype},
+    {outmasktype},
+    {opmasktype},
+    {bm},
+    {bn},
+    {bk},
+    {wm},
+    {wn},
+    {trans_a},
+    {trans_b},
+    {mn_aligned},
+    {k_aligned}>(
+    const device {itype}* A [[buffer(0)]],
+    const device {itype}* B [[buffer(1)]],
+    device {itype}* D [[buffer(3)]],
+    const constant GEMMParams* params [[buffer(4)]],
+    const constant int* batch_shape [[buffer(6)]],
+    const constant size_t* batch_strides [[buffer(7)]],
+    const device {outmasktype}* out_mask [[buffer(10)]],
+    const device {opmasktype}* lhs_mask [[buffer(11)]],
+    const device {opmasktype}* rhs_mask [[buffer(12)]],
+    const constant int* mask_strides [[buffer(13)]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint3 lid [[thread_position_in_threadgroup]]);
+)";
+
+constexpr std::string_view steel_gemm_splitk_kernels = R"(
+template [[host_name("{name}")]] [[kernel]] void
+gemm_splitk<
+    {itype},
+    {otype},
+    {bm},
+    {bn},
+    {bk},
+    {wm},
+    {wn},
+    {trans_a},
+    {trans_b},
+    {mn_aligned},
+    {k_aligned}>(
+    const device {itype}* A [[buffer(0)]],
+    const device {itype}* B [[buffer(1)]],
+    device {otype}* C [[buffer(2)]],
+    const constant GEMMSpiltKParams* params [[buffer(3)]],
+    uint simd_lane_id [[thread_index_in_simdgroup]],
+    uint simd_group_id [[simdgroup_index_in_threadgroup]],
+    uint3 tid [[threadgroup_position_in_grid]],
+    uint3 lid [[thread_position_in_threadgroup]]);
+)";
+
+constexpr std::string_view steel_gemm_splitk_accum_kernels = R"(
+template [[host_name("{name}")]] [[kernel]] void
+gemm_splitk_accum<{atype}, {otype}>(
+    const device {atype}* C_split [[buffer(0)]],
+    device {otype}* D [[buffer(1)]],
+    const constant int& k_partitions [[buffer(2)]],
+    const constant int& partition_stride [[buffer(3)]],
+    const constant int& ldd [[buffer(4)]],
+    uint2 gid [[thread_position_in_grid]]);
+)";
+
+constexpr std::string_view steel_gemm_splitk_accum_axbpy_kernels = R"(
+template [[host_name("{name}")]] [[kernel]] void
+gemm_splitk_accum_axpby<{atype}, {otype}>(
+    const device {atype}* C_split [[buffer(0)]],
+    device {otype}* D [[buffer(1)]],
+    const constant int& k_partitions [[buffer(2)]],
+    const constant int& partition_stride [[buffer(3)]],
+    const constant int& ldd [[buffer(4)]],
+    const device {otype}* C [[buffer(5)]],
+    const constant int& ldc [[buffer(6)]],
+    const constant int& fdc [[buffer(7)]],
+    const constant float& alpha [[buffer(8)]],
+    const constant float& beta [[buffer(9)]],
+    uint2 gid [[thread_position_in_grid]]);
+)";