Xnn s32 or #6823
Xnn s32 or #6823
Conversation
| xnn_storeu_s32(output + 1 * xnn_simd_size_s32, vy_1); | ||
| output += 32; | ||
| } | ||
| for (; batch >= xnn_simd_bytes_s32; batch -= xnn_simd_bytes_s32) { |
There was a problem hiding this comment.
you use batch -= 32 * sizeof(int32_t)) { for the main loop. which is the correct amount for avx512
should this one be batch -= 16 * sizeof(int32_t)) {
There was a problem hiding this comment.
@fbarchard vin1_0, vin1_1 two set of inputs are processed here, like loop unrolling of 2 so its 32
There was a problem hiding this comment.
vin1 is from input_a and is 16 ints
vin2 is from input_b and is 16 ints
should the loop be doing:
for (; batch >= 16 * sizeof(int32_t); batch -= 16 * sizeof(int32_t)) {
| output += xnn_simd_size_s32; | ||
| } | ||
| if XNN_UNLIKELY(batch != 0) { | ||
| xnn_simd_s32_t vin1 = xnn_load_tail_s32(input_a, batch >> XNN_LOG2_SIZEOF_INT32_T); |
There was a problem hiding this comment.
normally for native we'd create a mask based on the remainder and use it for all the instructions.
the loop here would be the same as the previous loop, but with a mask
its also prudent to put an assert on the expected batch size to ensure we dont accidently have too much or too little for the remainder masking to work
|
OR op is part of #6836. Hence closing it |
Bitwise OR implementation