Coverage for src/flag_gems/runtime/backend/_ascend/ops/max.py: 0%

1import logging 

2import math 

3from collections import namedtuple 

4 

5import torch 

6import triton 

7import triton.language as tl 

8 

9from flag_gems import runtime 

10from flag_gems.runtime import torch_device_fn 

11from flag_gems.utils import libentry 

12from flag_gems.utils import triton_lang_extension as tle 

13from flag_gems.utils.limits import get_dtype_min 

14 

15logger = logging.getLogger(f'flag_gems.runtime._ascend.ops.{__name__.split(".")[-1]}') 

16 

17 

18@libentry() 

19@triton.jit 

20def max_kernel_1( 

21 inp, 

22 mid, 

23 M, 

24 BLOCK_SIZE: tl.constexpr, 

25): 

26 pid = tle.program_id(0) 

27 offset = pid * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE) 

28 inp_ptrs = inp + offset 

29 mask = offset < M 

30 min_value = get_dtype_min(inp.type.element_ty) 

31 inp_val = tl.load(inp_ptrs, mask=mask, other=min_value) 

32 max_val = tl.max(inp_val) 

33 mid_ptr = mid + pid 

34 tl.store(mid_ptr, max_val) 

35 

36 

37@libentry() 

38@triton.jit 

39def max_kernel_2(mid, out, mid_size, BLOCK_MID: tl.constexpr): 

40 offset = tl.arange(0, BLOCK_MID) 

41 mid_ptrs = mid + offset 

42 mask = offset < mid_size 

43 min_value = get_dtype_min(mid.type.element_ty) 

44 mid_val = tl.load(mid_ptrs, mask=mask, other=min_value) 

45 max_val = tl.max(mid_val) 

46 tl.store(out, max_val) 

47 

48 

49def heur_block_n(args): 

50 return triton.next_power_of_2(args["N"]) 

51 

52 

53@libentry() 

54@triton.autotune( 

55 configs=runtime.get_tuned_config("max"), 

56 key=[ 

57 "M", 

58 "N", 

59 ], 

60) 

61@triton.jit 

62def max_kernel( 

63 inp, 

64 out_value, 

65 out_index, 

66 M, 

67 N, 

68 K, 

69 BLOCK_M: tl.constexpr, 

70 BLOCK_N: tl.constexpr, 

71): 

72 # set offset 

73 pid_m = tle.program_id(0) 

74 pid_k = tle.program_id(1) 

75 m_offset = pid_m * BLOCK_M + tl.arange(0, BLOCK_M) 

76 

77 dtype = inp.type.element_ty 

78 acc_type = tl.float32 if dtype is tl.bfloat16 else dtype 

79 min_value = get_dtype_min(dtype) 

80 result_value = tl.full([BLOCK_M], value=min_value, dtype=acc_type) 

81 result_index = tl.zeros([BLOCK_M], dtype=tl.int64) 

82 for i in range(0, N, BLOCK_N): 

83 n_offset = i + tl.arange(0, BLOCK_N) 

84 offset = m_offset[:, None] * N * K + n_offset[None, :] * K + pid_k 

85 # set mask 

86 mask = m_offset[:, None] < M and n_offset[None, :] < N 

87 inp_ptrs = inp + offset 

88 inp_vals = tl.load(inp_ptrs, mask=mask, other=min_value) 

89 if dtype is tl.int64: 

90 inp_vals = tl.where(mask, inp_vals, min_value) 

91 max_value, max_index = tl.max(inp_vals, axis=1, return_indices=True) 

92 update_mask = max_value > result_value 

93 result_value = tl.where(update_mask, max_value, result_value) 

94 result_index = tl.where(update_mask, i + max_index, result_index) 

95 mask1 = m_offset < M 

96 offset_index = m_offset * K + pid_k 

97 out_value_ptrs = out_value + offset_index 

98 out_index_ptrs = out_index + offset_index 

99 

100 tl.store(out_value_ptrs, result_value, mask=mask1) 

101 tl.store(out_index_ptrs, result_index, mask=mask1) 

102 

103 

104def max(inp): 

105 logger.debug("GEMS_ASCEND MAX") 

106 inp = inp.contiguous() 

107 M = inp.numel() 

108 block_size = triton.next_power_of_2(math.ceil(math.sqrt(M))) 

109 mid_size = triton.cdiv(M, block_size) 

110 block_mid = triton.next_power_of_2(mid_size) 

111 

112 dtype = inp.dtype 

113 mid = torch.empty((mid_size,), dtype=dtype, device=inp.device) 

114 out = torch.empty([], dtype=dtype, device=inp.device) 

115 

116 with torch_device_fn.device(inp.device): 

117 max_kernel_1[(mid_size, 1, 1)](inp, mid, M, block_size) 

118 max_kernel_2[(1, 1, 1)](mid, out, mid_size, block_mid) 

119 return out 

120 

121 

122def max_dim(inp, dim=None, keepdim=False): 

123 logger.debug("GEMS_ASCEND MAX") 

124 assert dim >= -inp.ndim and dim < inp.ndim, "Invalid dim" 

125 shape = inp.shape 

126 dim = dim % inp.ndim 

127 N = shape[dim] 

128 M = math.prod(shape[:dim]) 

129 K = inp.numel() // M // N 

130 

131 inp = inp.contiguous() 

132 

133 shape_list = list(shape) 

134 shape_list[dim] = 1 

135 out_value = torch.empty(shape_list, dtype=inp.dtype, device=inp.device) 

136 out_index = torch.empty(shape_list, dtype=torch.int64, device=inp.device) 

137 

138 if not keepdim: 

139 out_value = torch.squeeze(out_value, dim) 

140 out_index = torch.squeeze(out_index, dim) 

141 

142 def grid(meta): 

143 axis0 = triton.cdiv(M, meta["BLOCK_M"]) 

144 axis1 = K 

145 while axis0 * axis1 >= 65536: 

146 axis0 = axis0 // 2 

147 return ( 

148 axis0, 

149 axis1, 

150 ) 

151 

152 with torch_device_fn.device(inp.device): 

153 max_kernel[grid](inp, out_value, out_index, M, N, K) 

154 Max_out = namedtuple("max", ["values", "indices"]) 

155 out = Max_out(values=out_value, indices=out_index) 

156 return out