erasure_code: optimize AVX2 GFNI single vector dot product

Signed-off-by: Marcel Cornu <marcel.d.cornu@intel.com>

erasure_code/gf_vect_dot_prod_avx2_gfni.asm

@@ -48,16 +48,16 @@
  %define tmp   r11
  %define tmp2  r10
  %define tmp3  r12      ; must be saved and restored
- %define tmp4  r13	;must be saved and restored
 
+ %define stack_size  1*8
  %define func(x) x: endbranch
  %macro FUNC_SAVE 0
-	push	r12
+        sub	    rsp, stack_size
-	push	r13
+        mov     [rsp + 0*8], r12
  %endmacro
  %macro FUNC_RESTORE 0
-	pop	r13
+        mov     r12, [rsp + 0*8]
-	pop	r12
+        add     rsp, stack_size
  %endmacro
 %endif
 
@@ -68,30 +68,35 @@
  %define arg3   r9
 
  %define arg4   r12     ; must be saved, loaded and restored
- %define arg5   r13 		; must be saved and restored
+ %define arg5   r15     ; must be saved and restored
  %define tmp    r11
  %define tmp2   r10
- %define tmp3   r14
+ %define tmp3   r13     ; must be saved and restored
- %define tmp4   r15
+ %define stack_size  4*16 + 3*8     ; must be an odd multiple of 8
- %define stack_size  0*16 + 5*8		; must be an odd multiple of 8
  %define arg(x)      [rsp + stack_size + 8 + 8*x]
 
  %define func(x) proc_frame x
  %macro FUNC_SAVE 0
         alloc_stack stack_size
-	mov	[rsp + 0*8], r12
+        vmovdqa [rsp + 0*16], xmm6
-	mov	[rsp + 1*8], r13
+        vmovdqa [rsp + 1*16], xmm7
-	mov	[rsp + 2*8], r14
+        vmovdqa [rsp + 2*16], xmm8
-	mov	[rsp + 3*8], r15
+        vmovdqa [rsp + 3*16], xmm9
+        mov     [rsp + 4*16 + 0*8], r12
+        mov     [rsp + 4*16 + 1*8], r13
+        mov     [rsp + 4*16 + 2*8], r15
         end_prolog
         mov     arg4, arg(4)
  %endmacro
 
  %macro FUNC_RESTORE 0
-	mov	r12,  [rsp + 0*8]
+        vmovdqa xmm6, [rsp + 0*16]
-	mov	r13,  [rsp + 1*8]
+        vmovdqa xmm7, [rsp + 1*16]
-	mov	r14,  [rsp + 2*8]
+        vmovdqa xmm8, [rsp + 2*16]
-	mov	r15,  [rsp + 3*8]
+        vmovdqa xmm9, [rsp + 3*16]
+        mov     r12,  [rsp + 4*16 + 0*8]
+        mov     r13,  [rsp + 4*16 + 1*8]
+        mov     r15,  [rsp + 4*16 + 2*8]
         add     rsp, stack_size
  %endmacro
 %endif
@@ -106,7 +111,6 @@
 %define vec_i  tmp2
 %define pos    rax
 
-
 %ifndef EC_ALIGNED_ADDR
 ;;; Use Un-aligned load/store
  %define XLDR vmovdqu
@@ -122,17 +126,94 @@
  %endif
 %endif
 
-%define xgft1  ymm2
+%define x0l     ymm0
+%define x0h     ymm1
+%define x0x     ymm2
 
-%define x0     ymm0
+%define xp1l    ymm3
-%define xp1    ymm1
+%define xp1h    ymm4
+%define xp1x    ymm5
+
+%define xgft1   ymm6
+%define xgft2   ymm7
+%define xgft3   ymm8
+
+%define xtmp1   ymm9
+
+%define x0      x0l
+%define xp1     xp1l
+%define xp2     xp2l
+%define xp3     xp3l
 
 default rel
 [bits 64]
+
 section .text
 
 ;;
-;; Encodes 32 bytes of all "k" sources into 32 bytes (single parity disk)
+;; Encodes 96 bytes of all "k" sources into 96 bytes (single parity disk)
+;;
+%macro ENCODE_96B 0
+        vpxor   xp1l, xp1l, xp1l
+        vpxor   xp1h, xp1h, xp1h
+        vpxor   xp1x, xp1x, xp1x
+        mov     tmp, mul_array
+        xor     vec_i, vec_i
+
+%%next_vect:
+        ;; load next source vector
+        mov     ptr, [src + vec_i]
+        XLDR    x0l, [ptr + pos]
+        XLDR    x0h, [ptr + pos + 32]
+        XLDR    x0x, [ptr + pos + 64]
+        add     vec_i, 8
+
+        vbroadcastsd xgft1, [tmp]
+        add     tmp, 8
+
+        GF_MUL_XOR VEX, x0l, xgft1, xtmp1, xp1l
+        GF_MUL_XOR VEX, x0h, xgft1, xtmp1, xp1h
+        GF_MUL_XOR VEX, x0x, xgft1, xtmp1, xp1x
+
+        cmp     vec_i, vec
+        jl      %%next_vect
+
+        XSTR    [dest1 + pos], xp1l
+        XSTR    [dest1 + pos + 32], xp1h
+        XSTR    [dest1 + pos + 64], xp1x
+%endmacro
+
+;;
+;; Encodes 64 bytes of all "k" sources into 64 bytes (single parity disk)
+;;
+%macro ENCODE_64B 0
+        vpxor   xp1l, xp1l, xp1l
+        vpxor   xp1h, xp1h, xp1h
+        mov     tmp, mul_array
+        xor     vec_i, vec_i
+
+%%next_vect:
+        ;; load next source vector
+        mov     ptr, [src + vec_i]
+        XLDR    x0l, [ptr + pos]
+        XLDR    x0h, [ptr + pos + 32]
+        add     vec_i, 8
+
+        vbroadcastsd xgft1, [tmp]
+        add     tmp, 8
+
+        GF_MUL_XOR VEX, x0l, xgft1, xtmp1, xp1l
+        GF_MUL_XOR VEX, x0h, xgft1, xgft1, xp1h
+
+        cmp     vec_i, vec
+        jl      %%next_vect
+
+        XSTR    [dest1 + pos], xp1l
+        XSTR    [dest1 + pos + 32], xp1h
+%endmacro
+
+;;
+;; Encodes 32 bytes of all "k" sources into 32 bytes (single parity disks)
 ;;
 %macro ENCODE_32B 0
         vpxor   xp1, xp1, xp1
@@ -140,8 +221,9 @@ section .text
         xor     vec_i, vec_i
 
 %%next_vect:
+        ;; load next source vector
         mov     ptr, [src + vec_i]
-	XLDR	x0, [ptr + pos]		;Get next source vector (32 bytes)
+        XLDR    x0, [ptr + pos]
         add	    vec_i, 8
 
         vbroadcastsd xgft1, [tmp]
@@ -156,56 +238,76 @@ section .text
 %endmacro
 
 ;;
-;; Encodes less than 32 bytes of all "k" sources and updates single parity disk
+;; Encodes less than 32 bytes of all "k" sources into single parity disks
 ;;
 %macro ENCODE_LT_32B 1
 %define %%LEN   %1
 
         vpxor   xp1, xp1, xp1
-	mov	tmp, mul_array
         xor     vec_i, vec_i
 
 %%next_vect:
+        ; get next source vector
         mov     ptr, [src + vec_i]
-	simd_load_avx2 x0, ptr + pos, %%LEN, tmp3, tmp4 ;Get next source vector
+        simd_load_avx2 x0, ptr + pos, %%LEN, tmp, tmp3
         add     vec_i, 8
 
-        vbroadcastsd xgft1, [tmp]
+        vbroadcastsd xgft1, [mul_array]
-	add	tmp, 8
+        add     mul_array, 8
 
         GF_MUL_XOR VEX, x0, xgft1, xgft1, xp1
 
         cmp     vec_i, vec
         jl      %%next_vect
 
-	lea 	tmp, [dest1 + pos]
+        ;; Store updated encoded data
-	simd_store_avx2 tmp, xp1, %%LEN, tmp3, tmp4 ;Store updated encoded data
+        lea     ptr, [dest1 + pos]
+        simd_store_avx2 ptr, xp1, %%LEN, tmp, vec_i
 %endmacro
 
 align 16
 mk_global gf_vect_dot_prod_avx2_gfni, function
 func(gf_vect_dot_prod_avx2_gfni)
         FUNC_SAVE
+
         xor     pos, pos
-	shl	vec, 3		;vec *= 8. Make vec_i count by 8
+        shl     vec, 3      ;; vec *= 8. Make vec_i count by 8
 
+        cmp     len, 96
+        jl      .len_lt_96
+
+.loop96:
+        ENCODE_96B
+
+        add     pos, 96     ;; Loop on 96 bytes at a time first
+        sub     len, 96
+        cmp     len, 96
+        jge     .loop96
+
+.len_lt_96:
+        cmp     len, 64
+        jl      .len_lt_64
+
+        ENCODE_64B
+
+        add     pos, 64     ;; encode next 64 bytes
+        sub     len, 64
+
+.len_lt_64:
         cmp     len, 32
-        jb      .len_lt_32
+        jl      .len_lt_32
-
-.loop32:
 
         ENCODE_32B
 
-	add	pos, 32			;Loop on 32 bytes at a time
+        add     pos, 32     ;; encode next 32 bytes
         sub     len, 32
-	cmp	len, 32
-	jge	.loop32
 
 .len_lt_32:
         cmp     len, 0
         jle     .exit
 
-        ENCODE_LT_32B len
+        ENCODE_LT_32B len   ;; encode final bytes
+
 .exit:
         vzeroupper