コグノスケ

1993                                          2000
OPeNBooK  ---,    1996                   ,--> オープンブック（The Towerの版権を持つ）
             +--> オープンブック9003 ---+--> シノミクス
9003, inc ---'
1990

Debian TestingのICD loader

$ apt-cache search ocl-icd-libopencl1

ocl-icd-libopencl1 - Generic OpenCL ICD Loader

cl_platform_idとdispatch

// ocl-icd/ocl_icd_loader.c

static inline void _find_and_check_platforms(cl_uint num_icds) {
  cl_uint i;

...

    cl_platform_id *platforms = (cl_platform_id *) malloc( sizeof(cl_platform_id) * num_platforms);
    error = (*plt_fn_ptr)(num_platforms, platforms, NULL);

...

    for(j=0; j<num_platforms; j++) {
      debug(D_LOG, "Checking platform %i", j);
      struct platform_icd *p=&_picds[_num_picds];
      char *param_value=NULL;
      p->extension_suffix=NULL;
      p->vicd=&_icds[i];
      p->pid=platforms[j];    //★★pid = platform IDのことらしい

      /* If clGetPlatformInfo is not exported and we are here, it
       * means that OCL_ICD_ASSUME_ICD_EXTENSION. Si we try to take it
       * from the dispatch * table. If that fails too, we have to
       * bail.
       */
      if (plt_info_ptr == NULL) {
        plt_info_ptr = p->pid->dispatch->clGetPlatformInfo;    //★★dispatchメンバが存在することを前提としている

      
// ocl-icd/khronos-headers/CL/cl.h

typedef struct _cl_platform_id *    cl_platform_id;


// ocl-icd/(build-dir)/ocl_icd_loader_gen.h

struct _cl_platform_id { struct _cl_icd_dispatch *dispatch; };    //★★dispatch以外は特に規定がなさそう


// ocl-icd/(build-dir)/ocl_icd.h

struct _cl_icd_dispatch {
#ifdef CL_VERSION_1_0
  CL_API_ENTRY cl_int (CL_API_CALL*clGetPlatformIDs)(
    cl_uint          /* num_entries */,
    cl_platform_id * /* platforms */,
    cl_uint *        /* num_platforms */
  ) CL_API_SUFFIX__VERSION_1_0;
#else
  CL_API_ENTRY cl_int (CL_API_CALL* clUnknown0)(void);
#endif

#ifdef CL_VERSION_1_0
  CL_API_ENTRY cl_int (CL_API_CALL*
  clGetPlatformInfo)(
    cl_platform_id   /* platform */,
    cl_platform_info /* param_name */,
    size_t           /* param_value_size */,
    void *           /* param_value */,
    size_t *         /* param_value_size_ret */
  ) CL_API_SUFFIX__VERSION_1_0;
#else
  CL_API_ENTRY cl_int (CL_API_CALL* clUnknown1)(void);
#endif

...

//★★こんな調子で関数ポインタの定義が延々と続く

自動ベクトル化を有効にするコード

// gcc/config/riscv/riscv.c

/* Implement TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_MODES.  */

static unsigned int
riscv_autovectorize_vector_modes (vector_modes *modes, bool)
{
  if (TARGET_VECTOR)
    {
      modes->safe_push (V64SImode);
      modes->safe_push (V32SImode);
    }

  return 0;
}

...

#undef TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_MODES
#define TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_MODES riscv_autovectorize_vector_modes

テスト用のint配列をコピーする関数

void *cpy(void *dst, const void *src, int n)
{
	int *d = dst;
	const int *s = src;
	int i;

	for (i = 0; i < n / sizeof(*d); i++) {
		d[i] = s[i];
	}

	return dst;
}

自動ベクトル化の結果

//★★自動ベクトル化、あり

                d[i] = s[i];
   100b8:       1202e007                vlw.v   v0,(t0)
   100bc:       10038393                addi    t2,t2,256
   100c0:       f0038793                addi    a5,t2,-256
   100c4:       10028293                addi    t0,t0,256
   100c8:       0207e027                vsw.v   v0,(a5)
        for (i = 0; i < n / sizeof(*d); i++) {
   100cc:       fee296e3                bne     t0,a4,100b8 <cpy+0x44>
   100d0:       00661293                slli    t0,a2,0x6
   100d4:       02568963                beq     a3,t0,10106 <cpy+0x92>
   100d8:       959a                    add     a1,a1,t1
   100da:       932a                    add     t1,t1,a0
                d[i] = s[i];
   100dc:       0005a383                lw      t2,0(a1)


//★★自動ベクトル化、なし

                d[i] = s[i];
   10080:       0005a303                lw      t1,0(a1)
        for (i = 0; i < n / sizeof(*d); i++) {
   10084:       0591                    addi    a1,a1,4
   10086:       0291                    addi    t0,t0,4
                d[i] = s[i];
   10088:       fe62ae23                sw      t1,-4(t0)
        for (i = 0; i < n / sizeof(*d); i++) {
   1008c:       fe759ae3                bne     a1,t2,10080 <cpy+0xc>
        }

        return dst;
}
   10090:       8082                    ret

ベクトル演算の加算、減算、乗算、除算、論理演算（and, or, xor）の定義

;; gcc/config/riscv/riscv.md

(define_attr "vecmode" "unknown,V32SI,V64SI"
  (const_string "unknown"))

...

;; Iterator for hardware supported vector modes.
(define_mode_iterator ANYV [(V32SI "TARGET_VECTOR")
			    (V64SI "TARGET_VECTOR")])

...


;;★★加算

(define_insn "add<mode>3"
  [(set (match_operand:ANYV	       0 "register_operand" "=v")
	(plus:ANYV (match_operand:ANYV 1 "register_operand" " v")
		   (match_operand:ANYV 2 "arith_operand"    " v")))]
  "TARGET_VECTOR"
  "vadd.vvt%0,%1,%2"
  [(set_attr "type" "arith")
   (set_attr "vecmode" "<MODE>")])


;;★★減算

(define_insn "sub<mode>3"
  [(set (match_operand:ANYV		0 "register_operand" "=v")
	(minus:ANYV (match_operand:ANYV 1 "register_operand" " v")
		    (match_operand:ANYV 2 "arith_operand"    " v")))]
  "TARGET_VECTOR"
  "vsub.vvt%0,%1,%2"
  [(set_attr "type" "arith")
   (set_attr "vecmode" "<MODE>")])


;;★★乗算

(define_insn "mul<mode>3"
  [(set (match_operand:ANYV	       0 "register_operand" "=v")
	(mult:ANYV (match_operand:ANYV 1 "register_operand" " v")
		   (match_operand:ANYV 2 "arith_operand"    " v")))]
  "TARGET_VECTOR"
  "vmul.vvt%0,%1,%2"
  [(set_attr "type" "arith")
   (set_attr "vecmode" "<MODE>")])


;;★★除算

;; This code iterator allows unsigned and signed division to be generated
;; from the same template.
(define_code_iterator any_div [div udiv mod umod])

(define_insn "<optab><mode>3"
  [(set (match_operand:ANYV		  0 "register_operand" "=v")
	(any_div:ANYV (match_operand:ANYV 1 "register_operand" " v")
		      (match_operand:ANYV 2 "arith_operand"    " v")))]
  "TARGET_VECTOR"
  "v<insn>.vvt%0,%1,%2"
  [(set_attr "type" "arith")
   (set_attr "vecmode" "<MODE>")])


;;★★論理演算

;; This code iterator allows the three bitwise instructions to be generated
;; from the same template.
(define_code_iterator any_bitwise [and ior xor])

...

(define_insn "<optab><mode>3"
  [(set (match_operand:ANYV		      0 "register_operand" "=v")
	(any_bitwise:ANYV (match_operand:ANYV 1 "register_operand" "%v")
			  (match_operand:ANYV 2 "arith_operand"    " v")))]
  "TAREGET_VECTOR"
  "v<insn>.vvt%0,%1,%2"
  [(set_attr "type" "logical")
   (set_attr "vecmode" "<MODE>")])

ベクトル四則演算、論理演算のサンプルプログラム

typedef int __v64si __attribute__((__vector_size__(256)));

void test()
{
	__v64si v10, v11, v12, v13;0;

	__asm__ volatile ("vlw.v %0, %1\n" : "=&v"(v10) : "A"(b[10]));
	__asm__ volatile ("vlw.v %0, %1\n" : "=&v"(v11) : "A"(b[20]));
	__asm__ volatile ("vlw.v %0, %1\n" : "=&v"(v12) : "A"(b[30]));
	__asm__ volatile ("vlw.v %0, %1\n" : "=&v"(v13) : "A"(b[40]));

	v10 = v11 + v12;
	v11 &= v12 - v13;
	v12 |= v13 * v10;
	v13 ^= v10 / v11;

	__asm__ volatile ("vsw.v %1, %0\n" : "=A"(b[40]) : "v"(v10));
	__asm__ volatile ("vsw.v %1, %0\n" : "=A"(b[50]) : "v"(v11));
	__asm__ volatile ("vsw.v %1, %0\n" : "=A"(b[60]) : "v"(v12));
	__asm__ volatile ("vsw.v %1, %0\n" : "=A"(b[70]) : "v"(v13));
}

ベクトル四則演算、論理演算のサンプルプログラム（逆アセンブル）

$ riscv32-unknown-elf-gcc b.c -nostdlib -g -Og -march=rv32gcv -mabi=ilp32f

$ riscv32-unknown-elf-objdump -dS a.out

...

        __asm__ volatile ("vlw.v %0, %1\n" : "=&v"(v13) : "A"(b[40]));
   10092:       0a028793                addi    a5,t0,160
   10096:       1207e207                vlw.v   v4,(a5)

        v10 = v11 + v12;
   1009a:       022081d7                vadd.vv v3,v2,v1
        v11 &= v12 - v13;
   1009e:       0a120057                vsub.vv v0,v1,v4
   100a2:       26010057                vand.vv v0,v0,v2
        v12 |= v13 * v10;
   100a6:       9641a157                vmul.vv v2,v4,v3
   100aa:       2a208157                vor.vv  v2,v2,v1
        v13 ^= v10 / v11;
   100ae:       863020d7                vdiv.vv v1,v3,v0
   100b2:       2e1200d7                vxor.vv v1,v1,v4

        __asm__ volatile ("vsw.v %1, %0\n" : "=A"(b[40]) : "v"(v10));
   100b6:       0207e1a7                vsw.v   v3,(a5)

...