Merge pull request #5380 from quic/topic/sgemm_direct_sme1_alpha_beta

SME1 based direct kernel (with alpha and beta) for cblas_sgemm level 3

common_level3.h

@@ -52,6 +52,13 @@ void sgemm_direct(BLASLONG M, BLASLONG N, BLASLONG K,
 	float * B, BLASLONG strideB,
 	float * R, BLASLONG strideR);
 
+void sgemm_direct_alpha_beta(BLASLONG M, BLASLONG N, BLASLONG K,
+	float alpha,
+	float * A, BLASLONG strideA,
+	float * B, BLASLONG strideB,
+	float beta,
+	float * R, BLASLONG strideR);
+
 int sgemm_direct_performant(BLASLONG M, BLASLONG N, BLASLONG K);
 
 int shgemm_beta(BLASLONG, BLASLONG, BLASLONG, float,

common_param.h

@@ -256,6 +256,7 @@ int (*shgemm_otcopy   )(BLASLONG, BLASLONG, hfloat16 *, BLASLONG, hfloat16 *);
 #endif
 #ifdef ARCH_ARM64
   void (*sgemm_direct) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG , float *, BLASLONG , float * , BLASLONG);
+  void (*sgemm_direct_alpha_beta) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float, float * , BLASLONG);
 #endif
 
   

common_s.h

@@ -49,6 +49,7 @@
 
 #define SGEMM_DIRECT_PERFORMANT    sgemm_direct_performant
 #define SGEMM_DIRECT		sgemm_direct
+#define  SGEMM_DIRECT_ALPHA_BETA		sgemm_direct_alpha_beta
 
 #define	SGEMM_ONCOPY		sgemm_oncopy
 #define	SGEMM_OTCOPY		sgemm_otcopy
@@ -218,6 +219,7 @@
 #elif ARCH_ARM64
 #define SGEMM_DIRECT_PERFORMANT    sgemm_direct_performant
 #define  SGEMM_DIRECT		gotoblas -> sgemm_direct
+#define  SGEMM_DIRECT_ALPHA_BETA		gotoblas -> sgemm_direct_alpha_beta
 #endif
 
 #define	SGEMM_ONCOPY		gotoblas -> sgemm_oncopy

interface/gemm.c

@@ -441,6 +441,9 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
   if (beta == 0 && alpha == 1.0 && order == CblasRowMajor && TransA == CblasNoTrans && TransB == CblasNoTrans) {
 	SGEMM_DIRECT(m, n, k, a, lda, b, ldb, c, ldc);
 	return;
+  }else if (order == CblasRowMajor && TransA == CblasNoTrans && TransB == CblasNoTrans) {
+	SGEMM_DIRECT_ALPHA_BETA(m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
+	return;
   }
 #endif
 #endif

kernel/CMakeLists.txt

@@ -255,9 +255,11 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
 	  GenerateNamedObjects("${KERNELDIR}/${SGEMMDIRECTPERFORMANT}" "" "gemm_direct_performant" false "" "" false  SINGLE)
 	    elseif (ARM64)
 	  set (SGEMMDIRECTKERNEL sgemm_direct_arm64_sme1.c)
+	  set (SGEMMDIRECTKERNEL_ALPHA_BETA sgemm_direct_alpha_beta_arm64_sme1.c)
 	  set (SGEMMDIRECTSMEKERNEL sgemm_direct_sme1.S)
 	  set (SGEMMDIRECTPREKERNEL sgemm_direct_sme1_preprocess.S)
 	  GenerateNamedObjects("${KERNELDIR}/${SGEMMDIRECTKERNEL}" "" "gemm_direct" false "" "" false SINGLE)
+	  GenerateNamedObjects("${KERNELDIR}/${SGEMMDIRECTKERNEL_ALPHA_BETA}" "" "gemm_direct_alpha_beta" false "" "" false SINGLE)
    	  if (HAVE_SME)
 	  GenerateNamedObjects("${KERNELDIR}/${SGEMMDIRECTSMEKERNEL}" "" "gemm_direct_sme1" false "" "" false SINGLE)
 	  GenerateNamedObjects("${KERNELDIR}/${SGEMMDIRECTPREKERNEL}" "" "gemm_direct_sme1_preprocess" false "" "" false SINGLE)

kernel/Makefile.L3

@@ -132,6 +132,7 @@ ifeq ($(TARGET_CORE), ARMV9SME)
 HAVE_SME = 1
 endif
 SGEMMDIRECTKERNEL = sgemm_direct_arm64_sme1.c
+SGEMMDIRECTKERNEL_ALPHA_BETA = sgemm_direct_alpha_beta_arm64_sme1.c
 endif
 endif
 endif
@@ -208,7 +209,8 @@ SKERNELOBJS += \
 endif
 ifeq ($(ARCH), arm64)
 SKERNELOBJS += \
-	sgemm_direct$(TSUFFIX).$(SUFFIX)
+	sgemm_direct$(TSUFFIX).$(SUFFIX) \
+	sgemm_direct_alpha_beta$(TSUFFIX).$(SUFFIX)
 ifdef HAVE_SME
 SKERNELOBJS += \
 	sgemm_direct_sme1$(TSUFFIX).$(SUFFIX) \
@@ -969,6 +971,8 @@ endif
 ifeq ($(ARCH), arm64)
 $(KDIR)sgemm_direct$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMMDIRECTKERNEL)
 	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
+$(KDIR)sgemm_direct_alpha_beta$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMMDIRECTKERNEL_ALPHA_BETA)
+	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
 ifdef HAVE_SME
 $(KDIR)sgemm_direct_sme1$(TSUFFIX).$(SUFFIX) : 
 	$(CC) $(CFLAGS) -c $(KERNELDIR)/sgemm_direct_sme1.S -UDOUBLE -UCOMPLEX  -o $@

kernel/arm64/sgemm_direct_alpha_beta_arm64_sme1.c

@@ -0,0 +1,199 @@
+/*
+ Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
+ SPDX-License-Identifier: BSD-3-Clause-Clear
+*/
+
+#include "common.h"
+#include <stdlib.h>
+#include <inttypes.h>
+#include <math.h>
+#include "sme_abi.h"
+#if defined(HAVE_SME)
+
+#if defined(__ARM_FEATURE_SME) && defined(__clang__) && __clang_major__ >= 16
+#include <arm_sme.h>
+#endif
+
+/* Function prototypes */
+extern void sgemm_direct_sme1_preprocess(uint64_t nbr, uint64_t nbc,\
+                                  const float * restrict a, float *  a_mod) __asm__("sgemm_direct_sme1_preprocess");
+
+/* Function Definitions */
+static uint64_t sve_cntw() {
+    uint64_t cnt;
+    asm volatile(
+        "rdsvl  %[res], #1\n"
+        "lsr    %[res], %[res], #2\n"
+        : [res] "=r" (cnt) ::
+    );
+    return cnt;
+}
+
+#if defined(__ARM_FEATURE_SME) && defined(__ARM_FEATURE_LOCALLY_STREAMING) && defined(__clang__) && __clang_major__ >= 16
+// Outer product kernel.
+// Computes a 2SVL x 2SVL block of C, utilizing all four FP32 tiles of ZA.
+__attribute__((always_inline)) inline void
+kernel_2x2(const float *A, const float *B, float *C, size_t shared_dim,
+           size_t ldc, size_t block_rows, size_t block_cols, float alpha, float beta)
+    __arm_out("za") __arm_streaming {
+
+  const uint64_t svl = svcntw();
+  size_t ldb = ldc;
+  // Predicate set-up
+  svbool_t pg = svptrue_b32();
+  svbool_t pg_a_0 = svwhilelt_b32_u64(0, block_rows);
+  svbool_t pg_a_1 = svwhilelt_b32_u64(svl, block_rows);
+
+  svbool_t pg_b_0 = svwhilelt_b32_u64(0, block_cols);
+  svbool_t pg_b_1 = svwhilelt_b32_u64(svl, block_cols);
+
+#define pg_c_0 pg_b_0
+#define pg_c_1 pg_b_1
+
+  svzero_za();
+  svfloat32_t beta_vec = svdup_f32(beta);
+    // Load C to ZA
+  for (size_t i = 0; i < MIN(svl, block_rows); i++) {
+	svfloat32_t row_c_0 = svld1(pg_c_0, &C[i * ldc]);
+	row_c_0 = svmul_x(pg, beta_vec, row_c_0);
+    svwrite_hor_za32_f32_m(/*tile*/0, /*slice*/i, pg_c_0, row_c_0);
+    
+	svfloat32_t row_c_1 = svld1(pg_c_1, &C[i * ldc + svl]);
+	row_c_1 = svmul_x(pg, beta_vec, row_c_1);
+    svwrite_hor_za32_f32_m(/*tile*/1, /*slice*/i, pg_c_1, row_c_1);
+  }
+  for (size_t i = svl; i < block_rows; i++) {
+    svfloat32_t row_c_0 = svld1(pg_c_0, &C[i * ldc]);
+	row_c_0 = svmul_x(pg, beta_vec, row_c_0);
+	svwrite_hor_za32_f32_m(/*tile*/2, /*slice*/i, pg_c_0, row_c_0);
+
+    svfloat32_t row_c_1 = svld1(pg_c_1, &C[i * ldc + svl]);
+	row_c_1 = svmul_x(pg, beta_vec, row_c_1);
+    svwrite_hor_za32_f32_m(/*tile*/3, /*slice*/i, pg_c_1, row_c_1);
+  }
+  
+  svfloat32_t alpha_vec = svdup_f32(alpha);
+  // Iterate through shared dimension (K)
+  for (size_t k = 0; k < shared_dim; k++) {
+    // Load column of A
+    svfloat32_t col_a_0 = svld1(pg_a_0, &A[k * svl]);
+	col_a_0 = svmul_x(pg, alpha_vec, col_a_0);
+    svfloat32_t col_a_1 = svld1(pg_a_1, &A[(k + shared_dim) * svl]);
+    col_a_1 = svmul_x(pg, alpha_vec, col_a_1);
+    // Load row of B
+    svfloat32_t row_b_0 = svld1(pg_b_0, &B[k * ldb]);
+    svfloat32_t row_b_1 = svld1(pg_b_1, &B[k * ldb + svl]);
+    // Perform outer product
+    svmopa_za32_m(/*tile*/0, pg, pg, col_a_0, row_b_0);
+    svmopa_za32_m(/*tile*/1, pg, pg, col_a_0, row_b_1);
+    svmopa_za32_m(/*tile*/2, pg, pg, col_a_1, row_b_0);
+    svmopa_za32_m(/*tile*/3, pg, pg, col_a_1, row_b_1);
+  }
+
+  // Store to C from ZA
+  for (size_t i = 0; i < MIN(svl, block_rows); i++) {
+    svst1_hor_za32(/*tile*/0, /*slice*/i, pg_c_0, &C[i * ldc]);
+    svst1_hor_za32(/*tile*/1, /*slice*/i, pg_c_1, &C[i * ldc + svl]);
+  }
+  for (size_t i = svl; i < block_rows; i++) {
+    svst1_hor_za32(/*tile*/2, /*slice*/i, pg_c_0, &C[i * ldc]);
+    svst1_hor_za32(/*tile*/3, /*slice*/i, pg_c_1, &C[i * ldc + svl]);
+  }
+}
+
+__arm_new("za") __arm_locally_streaming
+void sgemm_direct_alpha_beta_sme1_2VLx2VL(uint64_t m, uint64_t k, uint64_t n, const float* alpha,\
+                                   const float *ba, const float *restrict bb, const float* beta,\
+                                   float *restrict C) {
+
+  const uint64_t num_rows = m;
+  const uint64_t num_cols = n;
+
+  const float *restrict a_ptr = ba;
+  const float *restrict b_ptr = bb;
+  float *restrict c_ptr = C;
+
+  const uint64_t svl = svcntw();
+  const uint64_t ldc = n;
+
+  // Block over rows of C (panels of A)
+  uint64_t row_idx = 0;
+
+  // 2x2 loop
+  uint64_t row_batch = 2*svl;
+
+  // Block over row dimension of C
+  for (; row_idx < num_rows; row_idx += row_batch) {
+    row_batch = MIN(row_batch, num_rows - row_idx);
+    uint64_t col_idx = 0;
+    uint64_t col_batch = 2*svl;
+
+    // Block over column dimension of C
+    for (; col_idx < num_cols; col_idx += col_batch) {
+      col_batch = MIN(col_batch, num_cols - col_idx);
+
+      kernel_2x2(&a_ptr[row_idx * k], &b_ptr[col_idx],
+                 &c_ptr[row_idx * ldc + col_idx], k,
+				 ldc, row_batch, col_batch, *alpha, *beta);
+    }
+  }
+  return;
+}
+
+#else
+void sgemm_direct_alpha_beta_sme1_2VLx2VL(uint64_t m, uint64_t k, uint64_t n, const float* alpha,\
+                                   const float *ba, const float *restrict bb, const float* beta,\
+                                   float *restrict C){}
+#endif
+
+/*void sgemm_kernel_direct (BLASLONG M, BLASLONG N, BLASLONG K,\
+       float * __restrict A, BLASLONG strideA, float * __restrict B,\
+       BLASLONG strideB , float * __restrict R, BLASLONG strideR)
+*/
+void CNAME (BLASLONG M, BLASLONG N, BLASLONG K, float alpha, float * __restrict A,\
+            BLASLONG strideA, float * __restrict B, BLASLONG strideB ,\
+            float beta, float * __restrict R, BLASLONG strideR){
+                
+        uint64_t m_mod, vl_elms;
+        
+        vl_elms = sve_cntw();
+
+        m_mod = ceil((double)M/(double)vl_elms) * vl_elms;
+
+        float *A_mod = (float *) malloc(m_mod*K*sizeof(float));
+	    
+	    /* Prevent compiler optimization by reading from memory instead
+	     * of reading directly from vector (z) registers.
+	     * */
+        asm volatile("" : : :"p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7",
+                         "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15",
+                         "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7",
+                         "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15",
+                         "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23",
+                         "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31");
+      
+        /* Pre-process the left matrix to make it suitable for 
+           matrix sum of outer-product calculation
+         */
+        sgemm_direct_sme1_preprocess(M, K, A, A_mod);
+       
+        asm volatile("" : : :"p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7",
+                         "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15",
+                         "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7",
+                         "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15",
+                         "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23",
+                         "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31");
+
+        /* Calculate C = alpha*A*B + beta*C */
+        sgemm_direct_alpha_beta_sme1_2VLx2VL(M, K, N, &alpha, A_mod, B, &beta, R);
+
+        free(A_mod);
+}
+
+#else
+
+void CNAME (BLASLONG M, BLASLONG N, BLASLONG K, float alpha, float * __restrict A,\
+            BLASLONG strideA, float * __restrict B, BLASLONG strideB ,\
+            float beta, float * __restrict R, BLASLONG strideR){}
+ 
+#endif

kernel/arm64/sme_abi.h

@@ -0,0 +1,46 @@
+/***************************************************************************
+ * Copyright (c) 2024, The OpenBLAS Project
+ * All rights reserved.
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ * 1. Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimer in
+ * the documentation and/or other materials provided with the
+ * distribution.
+ * 3. Neither the name of the OpenBLAS project nor the names of
+ * its contributors may be used to endorse or promote products
+ * derived from this software without specific prior written permission.
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+ * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ * ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+ * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+ * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE
+ * GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+ * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
+ * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF
+ * THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ * *****************************************************************************/
+
+#pragma once
+
+#include <stdlib.h>
+
+/**
+ *  * These are SME ABI routines for saving & restoring SME state.
+ *   * They are typically provided by a compiler runtime library such
+ *    * as libgcc or compiler-rt, but support for these routines is not
+ *     * yet available on all platforms.
+ *      *
+ *       * Define these as aborting stubs so that we loudly fail on nested
+ *        * usage of SME state.
+ *         *
+ *          * These are defined as weak symbols so that a compiler runtime can
+ *           * override them if supported.
+ *            */
+__attribute__((weak)) void __arm_tpidr2_save() { abort(); }
+__attribute__((weak)) void __arm_tpidr2_restore() { abort(); }
+

kernel/setparam-ref.c

@@ -215,6 +215,7 @@ gotoblas_t TABLE_NAME = {
 #endif
 #ifdef ARCH_ARM64
   sgemm_directTS,
+  sgemm_direct_alpha_betaTS,
 #endif
 
   sgemm_kernelTS, sgemm_betaTS,