aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp208
1 files changed, 104 insertions, 104 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp
index 1c1196b7a6..182d246d7a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2020, 2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -52,60 +52,60 @@ void a64_gemm_u8_8x12_x1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t
__asm __volatile (
// Initialize result registers, load initial operands, prime prefetches.
- "movi v8.4s, #0x0\n"
- "ldr %q[a0], [%[a_ptr]]\n"
- "movi v9.4s, #0x0\n"
- "ldr %q[b0], [%[b_ptr]]\n"
- "movi v10.4s, #0x0\n"
- "ldr %q[a1], [%[a_ptr], #16]\n"
- "movi v11.4s, #0x0\n"
- "ldr %q[b1], [%[b_ptr], #16]\n"
- "movi v12.4s, #0x0\n"
+ "movi v8.4s, #0x0\n"
+ "ldr %q[a0], [%[a_ptr]]\n"
+ "movi v9.4s, #0x0\n"
+ "ldr %q[b0], [%[b_ptr]]\n"
+ "movi v10.4s, #0x0\n"
+ "ldr %q[a1], [%[a_ptr], #16]\n"
+ "movi v11.4s, #0x0\n"
+ "ldr %q[b1], [%[b_ptr], #16]\n"
+ "movi v12.4s, #0x0\n"
ASM_PREFETCH("[%[b_ptr], #64]")
- "movi v13.4s, #0x0\n"
+ "movi v13.4s, #0x0\n"
ASM_PREFETCH("[%[a_ptr], #64]")
- "movi v14.4s, #0x0\n"
+ "movi v14.4s, #0x0\n"
ASM_PREFETCH("[%[b_ptr], #128]")
- "movi v15.4s, #0x0\n"
+ "movi v15.4s, #0x0\n"
ASM_PREFETCH("[%[a_ptr], #128]")
- "movi v16.4s, #0x0\n"
+ "movi v16.4s, #0x0\n"
ASM_PREFETCH("[%[b_ptr], #192]")
- "movi v17.4s, #0x0\n"
+ "movi v17.4s, #0x0\n"
ASM_PREFETCH("[%[b_ptr], #256]")
- "movi v18.4s, #0x0\n"
+ "movi v18.4s, #0x0\n"
ASM_PREFETCH("[%[a_ptr], #192]")
- "movi v19.4s, #0x0\n"
+ "movi v19.4s, #0x0\n"
ASM_PREFETCH("[%[b_ptr], #320]")
- "movi v20.4s, #0x0\n"
+ "movi v20.4s, #0x0\n"
ASM_PREFETCH("[%[a_ptr], #256]")
- "movi v21.4s, #0x0\n"
+ "movi v21.4s, #0x0\n"
ASM_PREFETCH("[%[b_ptr], #384]")
- "movi v22.4s, #0x0\n"
- "movi v23.4s, #0x0\n"
- "movi v24.4s, #0x0\n"
- "movi v25.4s, #0x0\n"
- "movi v26.4s, #0x0\n"
- "movi v27.4s, #0x0\n"
- "movi v28.4s, #0x0\n"
- "movi v29.4s, #0x0\n"
- "movi v30.4s, #0x0\n"
- "movi v31.4s, #0x0\n"
+ "movi v22.4s, #0x0\n"
+ "movi v23.4s, #0x0\n"
+ "movi v24.4s, #0x0\n"
+ "movi v25.4s, #0x0\n"
+ "movi v26.4s, #0x0\n"
+ "movi v27.4s, #0x0\n"
+ "movi v28.4s, #0x0\n"
+ "movi v29.4s, #0x0\n"
+ "movi v30.4s, #0x0\n"
+ "movi v31.4s, #0x0\n"
// Skip loop if we are doing zero iterations of it.
- "cbz %w[k], 4f\n"
+ "cbz %w[k], 4f\n"
// Loop proper
"1:\n"
".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
- "ldr %q[b2], [%[b_ptr], #32]\n"
+ "ldr %q[b2], [%[b_ptr], #32]\n"
".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
".word 0x6fa1e04d // udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
- "ldr %q[b0], [%[b_ptr], #48]\n"
+ "ldr %q[b0], [%[b_ptr], #48]\n"
".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
".word 0x6fa0e071 // udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
@@ -116,74 +116,74 @@ void a64_gemm_u8_8x12_x1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t
".word 0x6fa1e075 // udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
- "ldr %q[b1], [%[b_ptr], #64]\n"
+ "ldr %q[b1], [%[b_ptr], #64]\n"
".word 0x6f80e098 // udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
".word 0x6fa0e099 // udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
ASM_PREFETCH("[%[b_ptr], #448]")
".word 0x6f80e89a // udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
".word 0x6fa0e89b // udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
- "ldr %q[a0], [%[a_ptr], #32]\n"
+ "ldr %q[a0], [%[a_ptr], #32]\n"
".word 0x6f81e09c // udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
".word 0x6fa1e09d // udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
- "ldr %q[a1], [%[a_ptr], #48]\n"
+ "ldr %q[a1], [%[a_ptr], #48]\n"
".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
- "ldr %q[b2], [%[b_ptr], #80]\n"
+ "ldr %q[b2], [%[b_ptr], #80]\n"
".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
".word 0x6fa1e04d // udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
- "ldr %q[b0], [%[b_ptr], #96]\n"
+ "ldr %q[b0], [%[b_ptr], #96]\n"
".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
".word 0x6fa0e071 // udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
ASM_PREFETCH("[%[b_ptr], #512]")
".word 0x6f80e872 // udot v18.4s, %[b1].16b, %[a0].4b[2]\n"
".word 0x6fa0e873 // udot v19.4s, %[b1].16b, %[a0].4b[3]\n"
- "subs %w[k], %w[k], #1\n"
+ "subs %w[k], %w[k], #1\n"
".word 0x6f81e074 // udot v20.4s, %[b1].16b, %[a1].4b[0]\n"
".word 0x6fa1e075 // udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
- "ldr %q[b1], [%[b_ptr], #112]\n"
+ "ldr %q[b1], [%[b_ptr], #112]\n"
".word 0x6f80e098 // udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
".word 0x6fa0e099 // udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
- "add %[a_ptr], %[a_ptr], #64\n"
+ "add %[a_ptr], %[a_ptr], #64\n"
".word 0x6f80e89a // udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
".word 0x6fa0e89b // udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
- "ldr %q[a0], [%[a_ptr]]\n"
+ "ldr %q[a0], [%[a_ptr]]\n"
".word 0x6f81e09c // udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
- "add %[b_ptr], %[b_ptr], #96\n"
+ "add %[b_ptr], %[b_ptr], #96\n"
".word 0x6fa1e09d // udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
- "ldr %q[a1], [%[a_ptr], #16]\n"
- "bne 1b\n"
+ "ldr %q[a1], [%[a_ptr], #16]\n"
+ "bne 1b\n"
// Target to use when K is 1 or 2 (i.e. zero iterations of main loop)
"4:\n"
// Branch to alternative tail for odd K
- "cbnz %w[oddk], 2f\n"
+ "cbnz %w[oddk], 2f\n"
// Detached final iteration (even K)
".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
- "ldr %q[b2], [%[b_ptr], #32]\n"
+ "ldr %q[b2], [%[b_ptr], #32]\n"
".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
".word 0x6fa1e04d // udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
- "ldr %q[b0], [%[b_ptr], #48]\n"
+ "ldr %q[b0], [%[b_ptr], #48]\n"
".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
".word 0x6fa0e071 // udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
@@ -193,142 +193,142 @@ void a64_gemm_u8_8x12_x1(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t
".word 0x6fa1e075 // udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
- "ldr %q[b1], [%[b_ptr], #64]\n"
+ "ldr %q[b1], [%[b_ptr], #64]\n"
".word 0x6f80e098 // udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
".word 0x6fa0e099 // udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
- "add %[a_ptr], %[a_ptr], #64\n"
+ "add %[a_ptr], %[a_ptr], #64\n"
".word 0x6f80e89a // udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
".word 0x6fa0e89b // udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
- "ldr %q[a0], [%[a_ptr], #-32]\n"
+ "ldr %q[a0], [%[a_ptr], #-32]\n"
".word 0x6f81e09c // udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
".word 0x6fa1e09d // udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
- "ldr %q[a1], [%[a_ptr], #-16]\n"
+ "ldr %q[a1], [%[a_ptr], #-16]\n"
".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
- "ldr %q[b2], [%[b_ptr], #80]\n"
+ "ldr %q[b2], [%[b_ptr], #80]\n"
".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
- "add %[b_ptr], %[b_ptr], #96\n"
+ "add %[b_ptr], %[b_ptr], #96\n"
".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
- "str q8, [%[c_ptr], #0]\n"
+ "str q8, [%[c_ptr], #0]\n"
".word 0x6fa0e071 // udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
- "str q16, [%[c_ptr], #16]\n"
+ "str q16, [%[c_ptr], #16]\n"
".word 0x6f80e098 // udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
- "str q24, [%[c_ptr], #32]\n"
+ "str q24, [%[c_ptr], #32]\n"
".word 0x6fa0e099 // udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
- "str q9, [%[c_ptr], #48]\n"
+ "str q9, [%[c_ptr], #48]\n"
".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
- "str q17, [%[c_ptr], #64]\n"
+ "str q17, [%[c_ptr], #64]\n"
".word 0x6f80e872 // udot v18.4s, %[b1].16b, %[a0].4b[2]\n"
- "str q25, [%[c_ptr], #80]\n"
+ "str q25, [%[c_ptr], #80]\n"
".word 0x6f80e89a // udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
- "str q10, [%[c_ptr], #96]\n"
+ "str q10, [%[c_ptr], #96]\n"
".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
- "str q18, [%[c_ptr], #112]\n"
+ "str q18, [%[c_ptr], #112]\n"
".word 0x6fa0e873 // udot v19.4s, %[b1].16b, %[a0].4b[3]\n"
- "str q26, [%[c_ptr], #128]\n"
+ "str q26, [%[c_ptr], #128]\n"
".word 0x6fa0e89b // udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
- "str q11, [%[c_ptr], #144]\n"
+ "str q11, [%[c_ptr], #144]\n"
".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
- "str q19, [%[c_ptr], #160]\n"
+ "str q19, [%[c_ptr], #160]\n"
".word 0x6f81e074 // udot v20.4s, %[b1].16b, %[a1].4b[0]\n"
- "str q27, [%[c_ptr], #176]\n"
+ "str q27, [%[c_ptr], #176]\n"
".word 0x6f81e09c // udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
- "str q12, [%[c_ptr], #192]\n"
+ "str q12, [%[c_ptr], #192]\n"
".word 0x6fa1e04d // udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
- "str q20, [%[c_ptr], #208]\n"
+ "str q20, [%[c_ptr], #208]\n"
".word 0x6fa1e075 // udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
- "str q28, [%[c_ptr], #224]\n"
+ "str q28, [%[c_ptr], #224]\n"
".word 0x6fa1e09d // udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
- "str q13, [%[c_ptr], #240]\n"
+ "str q13, [%[c_ptr], #240]\n"
".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
- "str q21, [%[c_ptr], #256]\n"
+ "str q21, [%[c_ptr], #256]\n"
".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
- "str q29, [%[c_ptr], #272]\n"
+ "str q29, [%[c_ptr], #272]\n"
".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
- "str q14, [%[c_ptr], #288]\n"
+ "str q14, [%[c_ptr], #288]\n"
".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
- "str q22, [%[c_ptr], #304]\n"
+ "str q22, [%[c_ptr], #304]\n"
".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
- "str q30, [%[c_ptr], #320]\n"
+ "str q30, [%[c_ptr], #320]\n"
".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
- "str q15, [%[c_ptr], #336]\n"
+ "str q15, [%[c_ptr], #336]\n"
- "b 3f\n"
+ "b 3f\n"
// Detached final iteration (odd K)
"2:\n"
".word 0x6f80e048 // udot v8.4s , %[b0].16b, %[a0].4b[0]\n"
- "ldr %q[b2], [%[b_ptr], #32]\n"
+ "ldr %q[b2], [%[b_ptr], #32]\n"
".word 0x6f80e070 // udot v16.4s, %[b1].16b, %[a0].4b[0]\n"
".word 0x6fa0e049 // udot v9.4s , %[b0].16b, %[a0].4b[1]\n"
- "str q8, [%[c_ptr], #0]\n"
+ "str q8, [%[c_ptr], #0]\n"
".word 0x6fa0e071 // udot v17.4s, %[b1].16b, %[a0].4b[1]\n"
- "str q16, [%[c_ptr], #16]\n"
+ "str q16, [%[c_ptr], #16]\n"
".word 0x6f80e098 // udot v24.4s, %[b2].16b, %[a0].4b[0]\n"
- "add %[b_ptr], %[b_ptr], #48\n"
- "add %[a_ptr], %[a_ptr], #32\n"
- "str q24, [%[c_ptr], #32]\n"
+ "add %[b_ptr], %[b_ptr], #48\n"
+ "add %[a_ptr], %[a_ptr], #32\n"
+ "str q24, [%[c_ptr], #32]\n"
".word 0x6fa0e099 // udot v25.4s, %[b2].16b, %[a0].4b[1]\n"
- "str q9, [%[c_ptr], #48]\n"
+ "str q9, [%[c_ptr], #48]\n"
".word 0x6f80e84a // udot v10.4s, %[b0].16b, %[a0].4b[2]\n"
- "str q17, [%[c_ptr], #64]\n"
+ "str q17, [%[c_ptr], #64]\n"
".word 0x6f80e872 // udot v18.4s, %[b1].16b, %[a0].4b[2]\n"
- "str q25, [%[c_ptr], #80]\n"
+ "str q25, [%[c_ptr], #80]\n"
".word 0x6f80e89a // udot v26.4s, %[b2].16b, %[a0].4b[2]\n"
- "str q10, [%[c_ptr], #96]\n"
+ "str q10, [%[c_ptr], #96]\n"
".word 0x6fa0e84b // udot v11.4s, %[b0].16b, %[a0].4b[3]\n"
- "str q18, [%[c_ptr], #112]\n"
+ "str q18, [%[c_ptr], #112]\n"
".word 0x6fa0e873 // udot v19.4s, %[b1].16b, %[a0].4b[3]\n"
- "str q26, [%[c_ptr], #128]\n"
+ "str q26, [%[c_ptr], #128]\n"
".word 0x6fa0e89b // udot v27.4s, %[b2].16b, %[a0].4b[3]\n"
- "str q11, [%[c_ptr], #144]\n"
+ "str q11, [%[c_ptr], #144]\n"
".word 0x6f81e04c // udot v12.4s, %[b0].16b, %[a1].4b[0]\n"
- "str q19, [%[c_ptr], #160]\n"
+ "str q19, [%[c_ptr], #160]\n"
".word 0x6f81e074 // udot v20.4s, %[b1].16b, %[a1].4b[0]\n"
- "str q27, [%[c_ptr], #176]\n"
+ "str q27, [%[c_ptr], #176]\n"
".word 0x6f81e09c // udot v28.4s, %[b2].16b, %[a1].4b[0]\n"
- "str q12, [%[c_ptr], #192]\n"
+ "str q12, [%[c_ptr], #192]\n"
".word 0x6fa1e04d // udot v13.4s, %[b0].16b, %[a1].4b[1]\n"
- "str q20, [%[c_ptr], #208]\n"
+ "str q20, [%[c_ptr], #208]\n"
".word 0x6fa1e075 // udot v21.4s, %[b1].16b, %[a1].4b[1]\n"
- "str q28, [%[c_ptr], #224]\n"
+ "str q28, [%[c_ptr], #224]\n"
".word 0x6fa1e09d // udot v29.4s, %[b2].16b, %[a1].4b[1]\n"
- "str q13, [%[c_ptr], #240]\n"
+ "str q13, [%[c_ptr], #240]\n"
".word 0x6f81e84e // udot v14.4s, %[b0].16b, %[a1].4b[2]\n"
- "str q21, [%[c_ptr], #256]\n"
+ "str q21, [%[c_ptr], #256]\n"
".word 0x6f81e876 // udot v22.4s, %[b1].16b, %[a1].4b[2]\n"
- "str q29, [%[c_ptr], #272]\n"
+ "str q29, [%[c_ptr], #272]\n"
".word 0x6f81e89e // udot v30.4s, %[b2].16b, %[a1].4b[2]\n"
- "str q14, [%[c_ptr], #288]\n"
+ "str q14, [%[c_ptr], #288]\n"
".word 0x6fa1e84f // udot v15.4s, %[b0].16b, %[a1].4b[3]\n"
- "str q22, [%[c_ptr], #304]\n"
+ "str q22, [%[c_ptr], #304]\n"
".word 0x6fa1e877 // udot v23.4s, %[b1].16b, %[a1].4b[3]\n"
- "str q30, [%[c_ptr], #320]\n"
+ "str q30, [%[c_ptr], #320]\n"
".word 0x6fa1e89f // udot v31.4s, %[b2].16b, %[a1].4b[3]\n"
- "str q15, [%[c_ptr], #336]\n"
+ "str q15, [%[c_ptr], #336]\n"
// Common tail
"3:\n"
- "str q23, [%[c_ptr], #352]\n"
- "str q31, [%[c_ptr], #368]\n"
- "add %[c_ptr], %[c_ptr], #384\n"
+ "str q23, [%[c_ptr], #352]\n"
+ "str q31, [%[c_ptr], #368]\n"
+ "add %[c_ptr], %[c_ptr], #384\n"
:
[a_ptr] "+r" (a_ptr), [b_ptr] "+r" (b_ptr), [c_ptr] "+r" (c_ptr),