ckernel_normalization_checked.h File Reference

#include <stddef.h>

Go to the source code of this file.

Functions
int	gelu_rows_tanh_checked_f32 (const float *input, size_t input_elements, size_t input_stride, float *output, size_t output_elements, size_t output_stride, float *scratch, size_t scratch_elements, size_t rows, size_t channels)
int	layernorm_rows_checked_f32 (const float *input, size_t input_elements, size_t input_stride, const float *gamma, size_t gamma_elements, const float *beta, size_t beta_elements, float *output, size_t output_elements, size_t output_stride, float *scratch, size_t scratch_elements, size_t rows, size_t channels, float epsilon)
int	layernorm_rows_checked_workspace (size_t rows, size_t channels, size_t *elements)

Function Documentation

gelu_rows_tanh_checked_f32()

int gelu_rows_tanh_checked_f32	(	const float *	input,
		size_t	input_elements,
		size_t	input_stride,
		float *	output,
		size_t	output_elements,
		size_t	output_stride,
		float *	scratch,
		size_t	scratch_elements,
		size_t	rows,
		size_t	channels
	)

Definition at line 5 of file gelu_rows_tanh_checked.c.

                                                                           {
    size_t ni, no, ns;
    if (!ck_checked_span(rows, channels, input_stride, &ni) ||
        !ck_checked_span(rows, channels, output_stride, &no) ||
        !ck_checked_span(rows, channels, channels, &ns) ||
        input_elements < ni || output_elements < no || scratch_elements < ns) return -2;
    if (!ck_checked_region(input, ni) || !ck_checked_region(output, no) ||
        !ck_checked_region(scratch, ns) || ck_checked_overlap(input, ni, output, no) ||
        ck_checked_overlap(input, ni, scratch, ns) ||
        ck_checked_overlap(output, no, scratch, ns)) return -1;
    for (size_t r = 0; r < rows; ++r)
        if (!ck_checked_finite(input + r * input_stride, channels)) return -3;
    for (size_t r = 0; r < rows; ++r) {
        memcpy(scratch + r * channels, input + r * input_stride, channels * sizeof(float));
        /* Historical gelu_exact_inplace implements tanh GELU, matching gelu_new.
         * This preserves its arithmetic; an erf provider is a different contract. */
        gelu_exact_inplace(scratch + r * channels, channels);
    }
    if (!ck_checked_finite(scratch, ns)) return -4;
    for (size_t r = 0; r < rows; ++r)
        memcpy(output + r * output_stride, scratch + r * channels, channels * sizeof(float));
    return 0;
}

References ck_checked_finite(), ck_checked_overlap(), ck_checked_region(), ck_checked_span(), and gelu_exact_inplace().

layernorm_rows_checked_f32()

int layernorm_rows_checked_f32	(	const float *	input,
		size_t	input_elements,
		size_t	input_stride,
		const float *	gamma,
		size_t	gamma_elements,
		const float *	beta,
		size_t	beta_elements,
		float *	output,
		size_t	output_elements,
		size_t	output_stride,
		float *	scratch,
		size_t	scratch_elements,
		size_t	rows,
		size_t	channels,
		float	epsilon
	)

Definition at line 16 of file layernorm_rows_checked.c.

                                                                                          {
    size_t ni, no, ns;
    if (layernorm_rows_checked_workspace(rows, channels, &ns) ||
        !ck_checked_span(rows, channels, input_stride, &ni) ||
        !ck_checked_span(rows, channels, output_stride, &no) ||
        input_elements < ni || output_elements < no || gamma_elements < channels ||
        beta_elements < channels || scratch_elements < ns) return -2;
    if (!isfinite(epsilon) || epsilon <= 0 || !ck_checked_region(input, ni) ||
        !ck_checked_region(output, no) || !ck_checked_region(gamma, channels) ||
        !ck_checked_region(beta, channels) || !ck_checked_region(scratch, ns)) return -1;
    if (ck_checked_overlap(output, no, input, ni) ||
        ck_checked_overlap(output, no, gamma, channels) ||
        ck_checked_overlap(output, no, beta, channels) ||
        ck_checked_overlap(output, no, scratch, ns) ||
        ck_checked_overlap(scratch, ns, input, ni) ||
        ck_checked_overlap(scratch, ns, gamma, channels) ||
        ck_checked_overlap(scratch, ns, beta, channels)) return -1;
    if (!ck_checked_finite(gamma, channels) || !ck_checked_finite(beta, channels)) return -3;
    for (size_t r = 0; r < rows; ++r)
        if (!ck_checked_finite(input + r * input_stride, channels)) return -3;
    float *means = scratch + rows * channels;
    float *rstd = means + rows;
    for (size_t r = 0; r < rows; ++r) {
        /* One row avoids signed token*channel indexing in the legacy ABI.
         * Its existing reduction order is preserved; no arithmetic is duplicated. */
        layernorm_naive_serial_matched_precision(input + r * input_stride, gamma, beta,
            scratch + r * channels, means + r, rstd + r, 1, (int)channels, epsilon);
        if (!isfinite(means[r]) || !isfinite(rstd[r]) ||
            !ck_checked_finite(scratch + r * channels, channels)) return -4;
    }
    for (size_t r = 0; r < rows; ++r)
        memcpy(output + r * output_stride, scratch + r * channels, channels * sizeof(float));
    return 0;
}

References ck_checked_finite(), ck_checked_overlap(), ck_checked_region(), ck_checked_span(), layernorm_naive_serial_matched_precision(), and layernorm_rows_checked_workspace().

layernorm_rows_checked_workspace()

int layernorm_rows_checked_workspace	(	size_t	rows,
		size_t	channels,
		size_t *	elements
	)

Definition at line 7 of file layernorm_rows_checked.c.

                                                                                     {
    /* Keep the legacy provider's signed loop indices below addition overflow. */
    if (!elements || !rows || !channels || channels > (size_t)INT_MAX - 8 ||
        channels > SIZE_MAX - 2 || rows > SIZE_MAX / (channels + 2) ||
        rows * (channels + 2) > SIZE_MAX / sizeof(float)) return -2;
    *elements = rows * (channels + 2);
    return 0;
}

Referenced by layernorm_rows_checked_f32().