diff options
| author | Josh Blum <josh@joshknows.com> | 2012-07-02 15:17:03 -0700 |
|---|---|---|
| committer | Josh Blum <josh@joshknows.com> | 2012-08-30 01:24:54 -0700 |
| commit | 2105aad243e2da01af914b5307712db6f4489d9c (patch) | |
| tree | abca5cab32b789569ba3d6c38fcca08eb3347b61 /volk/include | |
| parent | 5892c29ee0509ff8aa226f14da524f7ec30f9654 (diff) | |
volk: replace num_bytes with num_points in many kernels
num_points is simply more intuitive than num_bytes
We expect to pass in the number of array elements.
This may have been the source of some test failures
when random numbers of points were interpreted as bytes,
you get caught slicing points/items down the middle...
Diffstat (limited to 'volk/include')
12 files changed, 77 insertions, 35 deletions
diff --git a/volk/include/volk/volk_16i_max_star_16i_a.h b/volk/include/volk/volk_16i_max_star_16i_a.h index edfff8a82b..ca81cf0d62 100644 --- a/volk/include/volk/volk_16i_max_star_16i_a.h +++ b/volk/include/volk/volk_16i_max_star_16i_a.h @@ -12,9 +12,9 @@ #include<emmintrin.h> #include<tmmintrin.h> -static inline void volk_16i_max_star_16i_a_ssse3(short* target, short* src0, unsigned int num_bytes) { - +static inline void volk_16i_max_star_16i_a_ssse3(short* target, short* src0, unsigned int num_points) { + const unsigned int num_bytes = num_points*2; short candidate = src0[0]; short cands[8]; @@ -87,7 +87,9 @@ static inline void volk_16i_max_star_16i_a_ssse3(short* target, short* src0, un #ifdef LV_HAVE_GENERIC -static inline void volk_16i_max_star_16i_a_generic(short* target, short* src0, unsigned int num_bytes) { +static inline void volk_16i_max_star_16i_a_generic(short* target, short* src0, unsigned int num_points) { + + const unsigned int num_bytes = num_points*2; int i = 0; diff --git a/volk/include/volk/volk_16i_max_star_horizontal_16i_a.h b/volk/include/volk/volk_16i_max_star_horizontal_16i_a.h index c1c9084256..13c235bc0b 100644 --- a/volk/include/volk/volk_16i_max_star_horizontal_16i_a.h +++ b/volk/include/volk/volk_16i_max_star_horizontal_16i_a.h @@ -13,7 +13,9 @@ #include<emmintrin.h> #include<tmmintrin.h> -static inline void volk_16i_max_star_horizontal_16i_a_ssse3(int16_t* target, int16_t* src0, unsigned int num_bytes) { +static inline void volk_16i_max_star_horizontal_16i_a_ssse3(int16_t* target, int16_t* src0, unsigned int num_points) { + + const unsigned int num_bytes = num_points*2; const static uint8_t shufmask0[16] = {0x00, 0x01, 0x04, 0x05, 0x08, 0x09, 0x0c, 0x0d, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff}; const static uint8_t shufmask1[16] = {0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x00, 0x01, 0x04, 0x05, 0x08, 0x09, 0x0c, 0x0d}; @@ -110,7 +112,9 @@ static inline void volk_16i_max_star_horizontal_16i_a_ssse3(int16_t* target, in #ifdef LV_HAVE_GENERIC -static inline void volk_16i_max_star_horizontal_16i_a_generic(int16_t* target, int16_t* src0, unsigned int num_bytes) { +static inline void volk_16i_max_star_horizontal_16i_a_generic(int16_t* target, int16_t* src0, unsigned int num_points) { + + const unsigned int num_bytes = num_points*2; int i = 0; diff --git a/volk/include/volk/volk_16i_permute_and_scalar_add_a.h b/volk/include/volk/volk_16i_permute_and_scalar_add_a.h index 47e3cbf9cb..d91b36208a 100644 --- a/volk/include/volk/volk_16i_permute_and_scalar_add_a.h +++ b/volk/include/volk/volk_16i_permute_and_scalar_add_a.h @@ -13,8 +13,9 @@ #include<xmmintrin.h> #include<emmintrin.h> -static inline void volk_16i_permute_and_scalar_add_a_sse2(short* target, short* src0, short* permute_indexes, short* cntl0, short* cntl1, short* cntl2, short* cntl3, short* scalars, unsigned int num_bytes) { +static inline void volk_16i_permute_and_scalar_add_a_sse2(short* target, short* src0, short* permute_indexes, short* cntl0, short* cntl1, short* cntl2, short* cntl3, short* scalars, unsigned int num_points) { + const unsigned int num_bytes = num_points*2; __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7; @@ -117,7 +118,9 @@ static inline void volk_16i_permute_and_scalar_add_a_sse2(short* target, short #ifdef LV_HAVE_GENERIC -static inline void volk_16i_permute_and_scalar_add_a_generic(short* target, short* src0, short* permute_indexes, short* cntl0, short* cntl1, short* cntl2, short* cntl3, short* scalars, unsigned int num_bytes) { +static inline void volk_16i_permute_and_scalar_add_a_generic(short* target, short* src0, short* permute_indexes, short* cntl0, short* cntl1, short* cntl2, short* cntl3, short* scalars, unsigned int num_points) { + + const unsigned int num_bytes = num_points*2; int i = 0; diff --git a/volk/include/volk/volk_16i_x4_quad_max_star_16i_a.h b/volk/include/volk/volk_16i_x4_quad_max_star_16i_a.h index 0d84985530..18b2e3d845 100644 --- a/volk/include/volk/volk_16i_x4_quad_max_star_16i_a.h +++ b/volk/include/volk/volk_16i_x4_quad_max_star_16i_a.h @@ -13,10 +13,9 @@ #include<emmintrin.h> -static inline void volk_16i_x4_quad_max_star_16i_a_sse2(short* target, short* src0, short* src1, short* src2, short* src3, unsigned int num_bytes) { - - +static inline void volk_16i_x4_quad_max_star_16i_a_sse2(short* target, short* src0, short* src1, short* src2, short* src3, unsigned int num_points) { + const unsigned int num_bytes = num_points*2; int i = 0; @@ -168,7 +167,9 @@ static inline void volk_16i_x4_quad_max_star_16i_a_sse2(short* target, short* s #ifdef LV_HAVE_GENERIC -static inline void volk_16i_x4_quad_max_star_16i_a_generic(short* target, short* src0, short* src1, short* src2, short* src3, unsigned int num_bytes) { +static inline void volk_16i_x4_quad_max_star_16i_a_generic(short* target, short* src0, short* src1, short* src2, short* src3, unsigned int num_points) { + + const unsigned int num_bytes = num_points*2; int i = 0; diff --git a/volk/include/volk/volk_16i_x5_add_quad_16i_x4_a.h b/volk/include/volk/volk_16i_x5_add_quad_16i_x4_a.h index 5560b92d92..677cb40e9f 100644 --- a/volk/include/volk/volk_16i_x5_add_quad_16i_x4_a.h +++ b/volk/include/volk/volk_16i_x5_add_quad_16i_x4_a.h @@ -13,7 +13,9 @@ #include<xmmintrin.h> #include<emmintrin.h> -static inline void volk_16i_x5_add_quad_16i_x4_a_sse2(short* target0, short* target1, short* target2, short* target3, short* src0, short* src1, short* src2, short* src3, short* src4, unsigned int num_bytes) { +static inline void volk_16i_x5_add_quad_16i_x4_a_sse2(short* target0, short* target1, short* target2, short* target3, short* src0, short* src1, short* src2, short* src3, short* src4, unsigned int num_points) { + + const unsigned int num_bytes = num_points*2; __m128i xmm0, xmm1, xmm2, xmm3, xmm4; __m128i *p_target0, *p_target1, *p_target2, *p_target3, *p_src0, *p_src1, *p_src2, *p_src3, *p_src4; @@ -113,7 +115,9 @@ static inline void volk_16i_x5_add_quad_16i_x4_a_sse2(short* target0, short* ta #ifdef LV_HAVE_GENERIC -static inline void volk_16i_x5_add_quad_16i_x4_a_generic(short* target0, short* target1, short* target2, short* target3, short* src0, short* src1, short* src2, short* src3, short* src4, unsigned int num_bytes) { +static inline void volk_16i_x5_add_quad_16i_x4_a_generic(short* target0, short* target1, short* target2, short* target3, short* src0, short* src1, short* src2, short* src3, short* src4, unsigned int num_points) { + + const unsigned int num_bytes = num_points*2; int i = 0; diff --git a/volk/include/volk/volk_32f_x3_sum_of_poly_32f_a.h b/volk/include/volk/volk_32f_x3_sum_of_poly_32f_a.h index 3c530628c8..e33e5a916a 100644 --- a/volk/include/volk/volk_32f_x3_sum_of_poly_32f_a.h +++ b/volk/include/volk/volk_32f_x3_sum_of_poly_32f_a.h @@ -13,8 +13,9 @@ #include<xmmintrin.h> #include<pmmintrin.h> -static inline void volk_32f_x3_sum_of_poly_32f_a_sse3(float* target, float* src0, float* center_point_array, float* cutoff, unsigned int num_bytes) { +static inline void volk_32f_x3_sum_of_poly_32f_a_sse3(float* target, float* src0, float* center_point_array, float* cutoff, unsigned int num_points) { + const unsigned int num_bytes = num_points*4; float result = 0.0; float fst = 0.0; @@ -100,9 +101,9 @@ static inline void volk_32f_x3_sum_of_poly_32f_a_sse3(float* target, float* src0 #ifdef LV_HAVE_GENERIC -static inline void volk_32f_x3_sum_of_poly_32f_a_generic(float* target, float* src0, float* center_point_array, float* cutoff, unsigned int num_bytes) { - +static inline void volk_32f_x3_sum_of_poly_32f_a_generic(float* target, float* src0, float* center_point_array, float* cutoff, unsigned int num_points) { + const unsigned int num_bytes = num_points*4; float result = 0.0; float fst = 0.0; diff --git a/volk/include/volk/volk_32fc_index_max_16u_a.h b/volk/include/volk/volk_32fc_index_max_16u_a.h index 842a6a0420..0e2201152c 100644 --- a/volk/include/volk/volk_32fc_index_max_16u_a.h +++ b/volk/include/volk/volk_32fc_index_max_16u_a.h @@ -11,9 +11,9 @@ #include<pmmintrin.h> -static inline void volk_32fc_index_max_16u_a_sse3(unsigned int* target, lv_32fc_t* src0, unsigned int num_bytes) { - +static inline void volk_32fc_index_max_16u_a_sse3(unsigned int* target, lv_32fc_t* src0, unsigned int num_points) { + const unsigned int num_bytes = num_points*8; union bit128 holderf; union bit128 holderi; @@ -189,7 +189,10 @@ static inline void volk_32fc_index_max_16u_a_sse3(unsigned int* target, lv_32fc_ #endif /*LV_HAVE_SSE3*/ #ifdef LV_HAVE_GENERIC -static inline void volk_32fc_index_max_16u_a_generic(unsigned int* target, lv_32fc_t* src0, unsigned int num_bytes) { +static inline void volk_32fc_index_max_16u_a_generic(unsigned int* target, lv_32fc_t* src0, unsigned int num_points) { + + const unsigned int num_bytes = num_points*8; + float sq_dist = 0.0; float max = 0.0; unsigned int index = 0; diff --git a/volk/include/volk/volk_32fc_x2_conjugate_dot_prod_32fc_a.h b/volk/include/volk/volk_32fc_x2_conjugate_dot_prod_32fc_a.h index e3dedf2fcd..0deb9c2f90 100644 --- a/volk/include/volk/volk_32fc_x2_conjugate_dot_prod_32fc_a.h +++ b/volk/include/volk/volk_32fc_x2_conjugate_dot_prod_32fc_a.h @@ -9,7 +9,9 @@ #ifdef LV_HAVE_GENERIC -static inline void volk_32fc_x2_conjugate_dot_prod_32fc_a_generic(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_bytes) { +static inline void volk_32fc_x2_conjugate_dot_prod_32fc_a_generic(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_points) { + + const unsigned int num_bytes = num_points*8; float * res = (float*) result; float * in = (float*) input; @@ -63,7 +65,9 @@ static inline void volk_32fc_x2_conjugate_dot_prod_32fc_a_generic(lv_32fc_t* res #if LV_HAVE_SSE && LV_HAVE_64 -static inline void volk_32fc_x2_conjugate_dot_prod_32fc_a_sse(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_bytes) { +static inline void volk_32fc_x2_conjugate_dot_prod_32fc_a_sse(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_points) { + + const unsigned int num_bytes = num_points*8; __VOLK_ATTR_ALIGNED(16) static const uint32_t conjugator[4]= {0x00000000, 0x80000000, 0x00000000, 0x80000000}; @@ -204,7 +208,9 @@ static inline void volk_32fc_x2_conjugate_dot_prod_32fc_a_sse(lv_32fc_t* result, #endif #if LV_HAVE_SSE && LV_HAVE_32 -static inline void volk_32fc_x2_conjugate_dot_prod_32fc_a_sse_32(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_bytes) { +static inline void volk_32fc_x2_conjugate_dot_prod_32fc_a_sse_32(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_points) { + + const unsigned int num_bytes = num_points*8; __VOLK_ATTR_ALIGNED(16) static const uint32_t conjugator[4]= {0x00000000, 0x80000000, 0x00000000, 0x80000000}; diff --git a/volk/include/volk/volk_32fc_x2_conjugate_dot_prod_32fc_u.h b/volk/include/volk/volk_32fc_x2_conjugate_dot_prod_32fc_u.h index e7493413f7..5b16b8639a 100644 --- a/volk/include/volk/volk_32fc_x2_conjugate_dot_prod_32fc_u.h +++ b/volk/include/volk/volk_32fc_x2_conjugate_dot_prod_32fc_u.h @@ -8,7 +8,9 @@ #ifdef LV_HAVE_GENERIC -static inline void volk_32fc_x2_conjugate_dot_prod_32fc_u_generic(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_bytes) { +static inline void volk_32fc_x2_conjugate_dot_prod_32fc_u_generic(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_points) { + + const unsigned int num_bytes = num_points*8; float * res = (float*) result; float * in = (float*) input; @@ -64,7 +66,9 @@ static inline void volk_32fc_x2_conjugate_dot_prod_32fc_u_generic(lv_32fc_t* res #include <mmintrin.h> -static inline void volk_32fc_x2_conjugate_dot_prod_32fc_u_sse3(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_bytes) { +static inline void volk_32fc_x2_conjugate_dot_prod_32fc_u_sse3(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_points) { + + unsigned int num_bytes = num_points*8; // Variable never used? //__VOLK_ATTR_ALIGNED(16) static const uint32_t conjugator[4]= {0x00000000, 0x80000000, 0x00000000, 0x80000000}; diff --git a/volk/include/volk/volk_32fc_x2_dot_prod_32fc_a.h b/volk/include/volk/volk_32fc_x2_dot_prod_32fc_a.h index caef3e6f0d..10ff4080ed 100644 --- a/volk/include/volk/volk_32fc_x2_dot_prod_32fc_a.h +++ b/volk/include/volk/volk_32fc_x2_dot_prod_32fc_a.h @@ -10,7 +10,9 @@ #ifdef LV_HAVE_GENERIC -static inline void volk_32fc_x2_dot_prod_32fc_a_generic(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_bytes) { +static inline void volk_32fc_x2_dot_prod_32fc_a_generic(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_points) { + + const unsigned int num_bytes = num_points*8; float * res = (float*) result; float * in = (float*) input; @@ -46,8 +48,9 @@ static inline void volk_32fc_x2_dot_prod_32fc_a_generic(lv_32fc_t* result, const #if LV_HAVE_SSE && LV_HAVE_64 -static inline void volk_32fc_x2_dot_prod_32fc_a_sse_64(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_bytes) { +static inline void volk_32fc_x2_dot_prod_32fc_a_sse_64(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_points) { + const unsigned int num_bytes = num_points*8; asm ( @@ -175,11 +178,11 @@ static inline void volk_32fc_x2_dot_prod_32fc_a_sse_64(lv_32fc_t* result, const #if LV_HAVE_SSE && LV_HAVE_32 -static inline void volk_32fc_x2_dot_prod_32fc_a_sse_32(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_bytes) { - - volk_32fc_x2_dot_prod_32fc_a_generic(result, input, taps, num_bytes); +static inline void volk_32fc_x2_dot_prod_32fc_a_sse_32(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_points) { + volk_32fc_x2_dot_prod_32fc_a_generic(result, input, taps, num_points); #if 0 + const unsigned int num_bytes = num_points*8; asm volatile ( " #pushl %%ebp\n\t" @@ -299,8 +302,9 @@ static inline void volk_32fc_x2_dot_prod_32fc_a_sse_32(lv_32fc_t* result, const #include <pmmintrin.h> -static inline void volk_32fc_x2_dot_prod_32fc_a_sse3(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_bytes) { +static inline void volk_32fc_x2_dot_prod_32fc_a_sse3(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_points) { + const unsigned int num_bytes = num_points*8; lv_32fc_t dotProduct; memset(&dotProduct, 0x0, 2*sizeof(float)); @@ -356,7 +360,9 @@ static inline void volk_32fc_x2_dot_prod_32fc_a_sse3(lv_32fc_t* result, const lv #include <smmintrin.h> -static inline void volk_32fc_x2_dot_prod_32fc_a_sse4_1(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_bytes) { +static inline void volk_32fc_x2_dot_prod_32fc_a_sse4_1(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_points) { + + const unsigned int num_bytes = num_points*8; __m128 xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7, real0, real1, im0, im1; float *p_input, *p_taps; diff --git a/volk/include/volk/volk_32fc_x2_s32f_square_dist_scalar_mult_32f_a.h b/volk/include/volk/volk_32fc_x2_s32f_square_dist_scalar_mult_32f_a.h index 75eb9173d5..d985fcd7f5 100644 --- a/volk/include/volk/volk_32fc_x2_s32f_square_dist_scalar_mult_32f_a.h +++ b/volk/include/volk/volk_32fc_x2_s32f_square_dist_scalar_mult_32f_a.h @@ -10,8 +10,9 @@ #include<xmmintrin.h> #include<pmmintrin.h> -static inline void volk_32fc_x2_s32f_square_dist_scalar_mult_32f_a_sse3(float* target, lv_32fc_t* src0, lv_32fc_t* points, float scalar, unsigned int num_bytes) { +static inline void volk_32fc_x2_s32f_square_dist_scalar_mult_32f_a_sse3(float* target, lv_32fc_t* src0, lv_32fc_t* points, float scalar, unsigned int num_points) { + const unsigned int num_bytes = num_points*8; __m128 xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7, xmm8; @@ -106,7 +107,10 @@ static inline void volk_32fc_x2_s32f_square_dist_scalar_mult_32f_a_sse3(float* t #endif /*LV_HAVE_SSE3*/ #ifdef LV_HAVE_GENERIC -static inline void volk_32fc_x2_s32f_square_dist_scalar_mult_32f_a_generic(float* target, lv_32fc_t* src0, lv_32fc_t* points, float scalar, unsigned int num_bytes) { +static inline void volk_32fc_x2_s32f_square_dist_scalar_mult_32f_a_generic(float* target, lv_32fc_t* src0, lv_32fc_t* points, float scalar, unsigned int num_points) { + + const unsigned int num_bytes = num_points*8; + lv_32fc_t diff; float sq_dist; unsigned int i = 0; diff --git a/volk/include/volk/volk_32fc_x2_square_dist_32f_a.h b/volk/include/volk/volk_32fc_x2_square_dist_32f_a.h index b819eaffd4..a10b6702bb 100644 --- a/volk/include/volk/volk_32fc_x2_square_dist_32f_a.h +++ b/volk/include/volk/volk_32fc_x2_square_dist_32f_a.h @@ -9,8 +9,9 @@ #include<xmmintrin.h> #include<pmmintrin.h> -static inline void volk_32fc_x2_square_dist_32f_a_sse3(float* target, lv_32fc_t* src0, lv_32fc_t* points, unsigned int num_bytes) { +static inline void volk_32fc_x2_square_dist_32f_a_sse3(float* target, lv_32fc_t* src0, lv_32fc_t* points, unsigned int num_points) { + const unsigned int num_bytes = num_points*8; __m128 xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7; @@ -92,7 +93,10 @@ static inline void volk_32fc_x2_square_dist_32f_a_sse3(float* target, lv_32fc_t* #endif /*LV_HAVE_SSE3*/ #ifdef LV_HAVE_GENERIC -static inline void volk_32fc_x2_square_dist_32f_a_generic(float* target, lv_32fc_t* src0, lv_32fc_t* points, unsigned int num_bytes) { +static inline void volk_32fc_x2_square_dist_32f_a_generic(float* target, lv_32fc_t* src0, lv_32fc_t* points, unsigned int num_points) { + + const unsigned int num_bytes = num_points*8; + lv_32fc_t diff; float sq_dist; unsigned int i = 0; |